Early warning control method and system for preventing sidewall wear-resistant rubber from being folded

Abstract

The invention relates to the technical field of tire manufacturing, and discloses an early warning control method and system for preventing sidewall wear-resistant rubber from being folded, the method is based on dynamic time warping and multi-modal data fusion analysis, multi-modal data such as sidewall wear-resistant rubber pressure distribution, temperature gradient and vibration signals are collected, and after denoising and feature extraction, the sidewall wear-resistant rubber is subjected to early warning control. And calculating the similarity with a standard deformation mode, and constructing a graph neural network model to evaluate the deformation risk probability. Self-adaptive early warning rules are generated through reinforcement learning, and whether early warning is conducted or not is monitored and judged in real time. And if the early warning is triggered, adjusting parameters of the sidewall forming machine, recording data, updating the standard deformation mode library, and iteratively optimizing the graph neural network model. The method can accurately monitor the risk, timely control and adjust, continuously optimize the model, effectively prevent the sidewall wear-resistant rubber from being folded, improve the tire production quality and efficiency, and reduce the production cost.

Description

Technical Field

[0001] The invention relates to the technical field of tire manufacturing, and in particular to an early warning control method and system for preventing sidewall wear-resistant rubber from being folded. Background Art

[0002] In the tire manufacturing industry, the quality of the sidewall wear-resistant rubber is directly related to the performance and service life of the tire. Discounting of the sidewall wear-resistant rubber is a common and thorny problem in the production process. It not only reduces the appearance quality of the tire, but also may cause safety hazards during the use of the tire, such as reduced sidewall strength and increased risk of tire blowout.

[0003] In the traditional tire manufacturing process, the quality control of the sidewall wear-resistant rubber mainly relies on manual experience and simple testing methods. Manual testing has the problems of low efficiency, strong subjectivity and difficulty in real-time monitoring. It is impossible to detect the slight changes of the sidewall wear-resistant rubber in the production process in time. By the time the discount problem is discovered, it has often caused a large amount of product waste and increased production costs.

[0004] With the continuous expansion of tire production scale and the increasing market requirements for tire quality, traditional quality control methods are increasingly unable to meet production needs. Some companies have tried to introduce automated testing equipment, but most of these devices can only detect a single physical quantity, such as pressure or temperature, and cannot fully and comprehensively evaluate the state of the sidewall wear-resistant rubber. Since the folding problem of the sidewall wear-resistant rubber is affected by a variety of factors, including uneven pressure distribution, abnormal temperature changes, and mechanical vibration, the detection of a single physical quantity cannot accurately predict and prevent the occurrence of folding.

[0005] In addition, the existing early warning system lacks flexibility in threshold setting, and usually adopts a fixed early warning threshold, which cannot be dynamically adjusted according to the actual situation in the production process. This leads to either a large number of false alarms in actual production, affecting production efficiency, or the failure to issue early warnings in time, causing problematic products to flow into the next process.

[0006] In terms of model optimization, traditional methods often ignore the in-depth use of historical data. The data in the tire production process has strong time series characteristics and correlations, but previous technologies have failed to fully tap the value of these data, making it impossible for the model to be continuously optimized as production data accumulates, and it is difficult to adapt to the complex and changing production environment. For example, when faced with slight changes in raw material properties, performance fluctuations caused by equipment aging, etc., the model cannot make timely adjustments, thereby reducing the accuracy of early warning and control of the sidewall wear rubber discount problem. Summary of the invention

[0007] The object of the present invention is to provide a warning control method and system for preventing the sidewall wear-resistant rubber from being folded so as to solve the problems raised in the above-mentioned background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: an early warning control method for preventing the sidewall wear-resistant rubber from being folded, the method comprising: S1. Collect multimodal data of the sidewall wear-resistant rubber during tire manufacturing, including pressure distribution, temperature gradient and vibration signal, mark the data with timestamps, and perform denoising and feature extraction on the original signal through wavelet transform to obtain a normalized feature matrix; S2. Divide the normalized feature matrix into continuous time segments, use the dynamic time warping algorithm to calculate the similarity between each time segment and the preset standard deformation pattern, and generate a deformation matching sequence; S3. Construct a graph neural network model of the sidewall structure, where nodes represent pressure monitoring points in the wear-resistant rubber area and edges represent pressure transmission relationships between adjacent areas; input the deformation matching sequence into the graph neural network and output the deformation risk probability of each node; S4. Based on the reinforcement learning framework, design an optimization strategy function, adjust the warning trigger threshold according to the historical deformation risk probability, and generate adaptive warning rules; S5. Real-time collection of multimodal data of the current production cycle, repeating steps S1 to S2, obtaining the current deformation matching sequence, inputting the graph neural network model of step S3, and obtaining the real-time deformation risk probability distribution; S6. Using a fuzzy logic system to process the real-time deformation risk probability distribution, calculate the comprehensive risk index, and determine whether to trigger a warning signal based on the adaptive warning rule of step S4; S7. If the warning signal is triggered, the pressure parameters and speed parameters of the sidewall molding machine are adjusted by the control execution unit, and the adjusted multimodal data is recorded to the historical database; S8. Reconstruct features of the adjusted data in the historical database using an autoencoder, calculate the reconstruction error vector, and update the standard deformation pattern library; S9. Based on the Bayesian probability model, combined with the real-time comprehensive risk index and historical reconstruction error, the weight parameters of the graph neural network are dynamically corrected to complete the iterative optimization of the model.

[0009] Preferably, the specific steps of the wavelet transform in step S1 include: S101. performing discrete wavelet decomposition on the pressure distribution signal to extract low-frequency approximate coefficients and high-frequency detail coefficients; S102. Performing a continuous wavelet transform on the temperature gradient signal to generate a time-frequency energy spectrum; S103. Perform wavelet packet decomposition on the vibration signal and select the optimal basis function to construct a frequency band energy distribution matrix; S104. Align the low-frequency approximation coefficients, the time-frequency energy spectrum, and the frequency band energy distribution matrix in time, merge them into a multidimensional feature matrix, and perform Z-score standardization.

[0010] Preferably, the calculation rule of the dynamic time warping algorithm in step S2 includes: by comparing the cumulative distance between the current time segment and the standard deformation mode on the time axis, selecting the optimal path of the previous moment for extension, wherein the cumulative distance is calculated by summing up the Euclidean distance between the corresponding points of the current time segment and the standard deformation mode; when the cumulative distance exceeds a preset deformation tolerance threshold, it is determined that there is a significant deviation between the time segment and the standard mode, and it is marked as a high-risk interval; the calculation formula is: ; in, Represents a time segment With standard mode The cumulative distance of is the Euclidean distance; when the cumulative distance exceeds the deformation tolerance threshold , mark the time segment as a high-risk interval.

[0011] Preferably, the update rules of the graph neural network model in step S3 include: concatenating the current layer embedding vector of each node with the aggregation vector of its neighboring nodes, performing a linear transformation through a trainable weight matrix, and then applying a nonlinear activation function to generate the node embedding vector of the next layer, wherein the aggregation of neighboring nodes is achieved through a summation operation.

[0012] Preferably, the reinforcement learning framework in step S4 adopts a deep Q network (DQN), whose state space is defined as the statistical characteristics of the deformation risk probability, the action space is the incremental adjustment of the warning threshold, and the reward function is designed as: ; in, Reward value, is the weight coefficient, and the network parameters are updated through the experience replay mechanism.

[0013] Preferably, the input variables of the fuzzy logic system in step S6 include real-time deformation risk probability and historical average risk , the output is the comprehensive risk index , and its fuzzy rule base is defined as: like High and High, then Very high; like Zhongqie Low, then medium; like Low and If (risk change rate) is high, then high, is the risk change rate.

[0014] Preferably, the reconstruction error calculation of the autoencoder in step S8 includes: calculating the square error between the input feature matrix and the reconstructed output, and adding a Frobenius norm regularization term of the encoder weight matrix to optimize the autoencoder parameters by minimizing the total error.

[0015] Preferably, it is characterized in that the updating process of the Bayesian probability model in step S9 includes: according to the observation data set consisting of the real-time comprehensive risk index and the historical reconstruction error, combined with the Gaussian prior distribution, the posterior probability distribution of the weight parameter is calculated by the likelihood function to complete the iterative correction of the graph neural network parameters.

[0016] Preferably, the parameter adjustment strategy of the control execution unit in step S7 includes: generating a proportional adjustment amount according to the difference between the comprehensive risk index and the dynamic warning threshold, generating a differential adjustment amount in combination with the real-time change rate of the risk index, and outputting the adjustment amount of the pressure parameter after weighted summation, and the calculation formula is: ; in, is the pressure adjustment amount, is a real-time comprehensive risk index. are the proportional and differential coefficients, is the dynamic warning threshold, is the rate of change of the comprehensive risk index.

[0017] Preferably, the present invention further includes an early warning control system for preventing the sidewall wear-resistant rubber from being folded, the system comprising: Data acquisition and preprocessing module: used to collect multimodal data of the sidewall wear-resistant rubber during tire manufacturing, including pressure distribution, temperature gradient and vibration signal, mark the data with timestamps, and perform denoising and feature extraction on the original signal through wavelet transform to obtain a normalized feature matrix; Deformation matching module: divides the normalized feature matrix into continuous time segments, uses the dynamic time warping algorithm to calculate the similarity between each time segment and the preset standard deformation pattern, and generates a deformation matching sequence; Graph neural network risk assessment module: Build a graph neural network model of the sidewall structure, where nodes represent pressure monitoring points in the wear-resistant rubber area, edges represent pressure transmission relationships in adjacent areas, receive deformation matching sequence input, and output deformation risk probability of each node; Warning rule generation module: Based on the reinforcement learning framework, the optimization strategy function is designed, the warning trigger threshold is adjusted according to the historical deformation risk probability, and adaptive warning rules are generated; Real-time monitoring and early warning judgment module: collect multimodal data of the current production cycle in real time, use the data acquisition and preprocessing module and the deformation matching module to obtain the current deformation matching sequence, input the graph neural network risk assessment module to obtain the real-time deformation risk probability distribution, use the fuzzy logic system to process the real-time deformation risk probability distribution, calculate the comprehensive risk index, and determine whether to trigger the early warning signal based on the adaptive early warning rules; Control execution module: if the warning signal is triggered, the pressure parameters and speed parameters of the sidewall forming machine are adjusted, and the adjusted multi-modal data are recorded in the historical database; Model iteration optimization module: Use autoencoders to reconstruct features of the adjusted data in the historical database, calculate the reconstruction error vector, update the standard deformation pattern library, and dynamically correct the weight parameters of the graph neural network based on the Bayesian probability model, combined with the real-time comprehensive risk index and historical reconstruction error.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The present invention collects multimodal data such as pressure distribution, temperature gradient and vibration signal of the side wear rubber, and uses wavelet transform for denoising and feature extraction, so as to comprehensively and accurately obtain the status information of the side wear rubber in the production process. By using the dynamic time warping algorithm to calculate the similarity with the preset standard deformation pattern, and constructing a graph neural network model to evaluate the deformation risk probability of each node, the accurate monitoring and evaluation of the discount risk of the side wear rubber is achieved, which greatly improves the accuracy of the early warning and reduces the false alarms and missed alarms. For example, in actual production, through the analysis of multimodal data, the potential discount risk caused by abnormal pressure in the local area of ​​the side wear rubber can be discovered in advance, providing a basis for taking timely measures.

[0019] Based on the optimization strategy function designed in the reinforcement learning framework, the warning trigger threshold is dynamically adjusted according to the historical deformation risk probability to generate an adaptive warning rule. This enables the warning system to better adapt to different production conditions and environmental changes, and improves the timeliness and effectiveness of the warning. Compared with the traditional fixed threshold warning method, the adaptive warning rule of the present invention can be flexibly adjusted according to the actual situation in the production process, avoiding the problem of inaccurate warning caused by unreasonable thresholds. For example, when a minor fault occurs in the production equipment, resulting in an increased risk of deformation of the sidewall wear-resistant rubber, the adaptive warning rule can timely improve the sensitivity of the warning, ensure that an alarm is issued at the beginning of the risk, and remind the operator to take corresponding measures. Once the warning signal is triggered, the control execution unit can quickly adjust the pressure parameters and speed parameters of the sidewall molding machine, effectively correct the abnormal deformation of the sidewall wear-resistant rubber, and prevent the occurrence of discounting. At the same time, the adjusted multimodal data is recorded in the historical database to provide data support for subsequent analysis and optimization. This not only improves the production quality of tires, reduces the defective rate, but also reduces production costs.

[0020] The autoencoder is used to reconstruct the features of the adjusted data in the historical database, calculate the reconstruction error vector, update the standard deformation pattern library, and dynamically correct the weight parameters of the graph neural network based on the Bayesian probability model combined with the real-time comprehensive risk index and the historical reconstruction error. This continuous optimization mechanism enables the system to continuously learn and adapt to changes in the production process. With the accumulation of production data, the accuracy and reliability of the model continue to improve, and the ability to prevent and control the sidewall wear rubber discount problem is also getting stronger. For example, with the replacement of raw material suppliers or the fine-tuning of production processes, the system can automatically adjust the model parameters through learning and analyzing new data to maintain effective control of the quality of the sidewall wear rubber. Through accurate early warning and timely control, the number of production interruptions and product rework caused by the sidewall wear rubber discount is reduced, and production efficiency is improved. At the same time, the adaptive early warning rules and continuously optimized models reduce the frequency and cost of manual intervention, making the production process more automated and intelligent. This advantage is particularly obvious in large-scale tire production, which can effectively shorten the production cycle and improve the production efficiency of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a working principle diagram of the early warning control method for preventing the sidewall wear-resistant rubber from being folded according to the present invention; Figure 2 Updated working principle diagram for graph neural network model; Figure 3 Reconstruct the working principle diagram of error calculation and optimization for the autoencoder; Figure 4 Flow chart for adjusting pressure parameters of control execution unit. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] See also Figure 1-4 The present invention provides a technical solution: an early warning control method for preventing the sidewall wear-resistant rubber from being folded, the method comprising: Data acquisition and preprocessing (S1): During the tire manufacturing process, various sensors are used to collect multimodal data of the tire sidewall wear-resistant rubber, including pressure distribution, temperature gradient and vibration signals. To ensure the time accuracy of the data, the collected data is timestamped. Subsequently, the original signal is denoised and feature extracted by wavelet transform. In actual operation, professional signal acquisition equipment such as high-precision pressure sensors, infrared temperature sensors and vibration acceleration sensors can be used to obtain the corresponding physical quantity data. The acquisition frequency can be set according to the actual needs of the tire manufacturing process, such as 100 times per second, to ensure that the changes in the data can be accurately captured.

[0024] Deformation matching calculation (S2): The normalized feature matrix obtained after preprocessing is divided into continuous time segments. The dynamic time warping algorithm is used to calculate the similarity between each time segment and the preset standard deformation pattern, thereby generating a deformation matching sequence. In actual application scenarios, the preset standard deformation pattern is obtained by analyzing and summarizing a large amount of sidewall wear-resistant rubber data under normal production conditions. This standard pattern represents the deformation law of the sidewall wear-resistant rubber during normal production, and provides an important reference for subsequent risk assessment.

[0025] Constructing a graph neural network model (S3): Constructing a graph neural network model of the sidewall structure, where nodes represent pressure monitoring points in the wear-resistant rubber area and edges represent pressure transfer relationships between adjacent areas. Input the deformation matching sequence into the graph neural network and output the deformation risk probability of each node. When constructing the graph neural network model, the topology of the model needs to be determined based on the actual structure of the sidewall and the distribution of pressure monitoring points. For example, for large tires, due to the large sidewall area, the distribution of pressure monitoring points is relatively sparse. When constructing the model, it is necessary to appropriately adjust the connection method of nodes and edges to ensure that the model can accurately reflect the pressure transfer relationship.

[0026] Design optimization strategy function (S4): Based on the reinforcement learning framework, design the optimization strategy function. Adjust the warning trigger threshold according to the historical deformation risk probability and generate adaptive warning rules. In practical applications, the reinforcement learning framework can use the deep Q network (DQN). Through continuous learning and optimization, the early warning system can dynamically adjust the warning threshold according to the actual production situation, improving the accuracy and timeliness of the early warning.

[0027] Real-time data acquisition and processing (S5): Real-time acquisition of multimodal data of the current production cycle, repeating steps S1 to S2 to obtain the current deformation matching sequence. Input the sequence into the graph neural network model of step S3 to obtain the real-time deformation risk probability distribution. In the real-time acquisition process, the same sensor equipment as step S1 is used to collect data at the same acquisition frequency. At the same time, efficient data transmission and processing technology is used to ensure that the current deformation matching sequence can be quickly acquired and input into the graph neural network model for risk assessment.

[0028] Calculation of comprehensive risk index and early warning judgment (S6): A fuzzy logic system is used to process the real-time deformation risk probability distribution and calculate the comprehensive risk index. Based on the adaptive early warning rule of step S4, it is determined whether to trigger the early warning signal. The input variables of the fuzzy logic system include the real-time deformation risk probability and the historical average risk. In practical applications, the fuzzy rule base is determined by analyzing a large amount of historical data and summarizing expert experience. For example, when the real-time deformation risk probability is high and the historical average risk is high, the comprehensive risk index is judged to be extremely high, triggering the early warning signal.

[0029] Parameter adjustment and data recording (S7): If the warning signal is triggered, the pressure parameters and speed parameters of the sidewall forming machine are adjusted through the control execution unit, and the adjusted multimodal data are recorded in the historical database. When adjusting the parameters, the proportional adjustment amount can be generated according to the difference between the comprehensive risk index and the dynamic warning threshold, and the differential adjustment amount can be generated in combination with the real-time change rate of the risk index. The adjustment amount of the pressure parameter is output after weighted summation. For example, when the comprehensive risk index is much higher than the dynamic warning threshold and the change rate is large, the adjustment range of the pressure parameter is increased to quickly correct the abnormal deformation of the sidewall wear-resistant rubber.

[0030] Standard deformation pattern library update (S8): The adjusted data in the historical database is reconstructed using an autoencoder, the reconstruction error vector is calculated, and the standard deformation pattern library is updated. In actual operation, the autoencoder optimizes its own parameters by minimizing the square error between the input feature matrix and the reconstructed output and adding the Frobenius norm regularization term of the encoder weight matrix. By continuously updating the standard deformation pattern library, it can better adapt to changes in the production process and improve the accuracy and reliability of the early warning system.

[0031] Iterative optimization of graph neural network model (S9): Based on the Bayesian probability model, combined with the real-time comprehensive risk index and historical reconstruction error, the weight parameters of the graph neural network are dynamically corrected to complete the iterative optimization of the model. In practical applications, based on the observation data set composed of the real-time comprehensive risk index and the historical reconstruction error, combined with the Gaussian prior distribution, the posterior probability distribution of the weight parameters is calculated through the likelihood function, thereby realizing the iterative correction of the graph neural network parameters. After multiple iterative optimizations, the graph neural network model can more accurately assess the deformation risk of the sidewall wear-resistant rubber, providing a more reliable guarantee to prevent the sidewall wear-resistant rubber from being discounted.

[0032] The present invention will be further described below in conjunction with Examples 1 to 5: Example

[0033] In the data acquisition stage, the pressure distribution signal collected by the pressure sensor has various noise interferences, such as environmental electromagnetic interference, sensor errors, etc. In order to remove these noises and extract useful features, the pressure distribution signal is subjected to discrete wavelet decomposition (S101). Specifically, a suitable wavelet basis function is selected, such as the commonly used Daubechies wavelet. According to the frequency characteristics of the signal, the number of decomposition layers is determined, assuming that it is decomposed into 3 layers. Through discrete wavelet decomposition, the pressure distribution signal is decomposed into low-frequency approximate coefficients and high-frequency detail coefficients. The low-frequency approximate coefficients retain the main trends and characteristics of the signal, while the high-frequency detail coefficients contain the signal's detailed information and noise. In practical applications, the high-frequency detail coefficients can be processed by setting a threshold to remove the noise components therein, and then reconstructed to obtain the denoised pressure distribution signal.

[0034] For the temperature gradient signal, since it has certain variation characteristics in time and frequency, the continuous wavelet transform (S102) can better analyze its energy distribution at different times and frequencies. Select Morlet wavelet as the basis function of continuous wavelet transform, transform the temperature gradient signal, and generate a time-frequency energy spectrum. The time-frequency energy spectrum can intuitively display the energy distribution of the temperature gradient signal at different times and frequencies, helping analysts to better understand the characteristics of the signal. For example, the time-frequency energy spectrum can find that the temperature gradient fluctuates abnormally in certain time periods, which may be related to the production process or equipment failure of the tire side wear-resistant rubber.

[0035] The frequency components of the vibration signal are relatively complex, and the wavelet packet decomposition (S103) can be used to analyze its frequency band energy distribution in more detail. In the wavelet packet decomposition process, the optimal basis function is selected according to the frequency characteristics of the signal to construct the frequency band energy distribution matrix. For example, by calculating indicators such as information entropy under different basis functions, the basis function with the smallest information entropy is selected as the optimal basis function. This ensures that the frequency band energy distribution matrix can more accurately reflect the characteristics of the vibration signal.

[0036] Finally, the low-frequency approximation coefficients, time-frequency energy spectrum and frequency band energy distribution matrix are aligned in time (S104). Since these data are obtained in different processing steps, their time scales may be different, so time alignment is required. Time alignment can be achieved through methods such as timestamp labeling and interpolation. Then, these data are merged into a multidimensional feature matrix and Z-score normalization is performed. Z-score normalization can map data with different characteristics to the same scale, eliminate the influence of data dimension, and make subsequent analysis and processing more accurate. For example, for the low-frequency approximation coefficients of the pressure distribution signal, the time-frequency energy spectrum of the temperature gradient signal, and the frequency band energy distribution matrix of the vibration signal, after Z-score normalization, their values ​​are all in the same range, which is convenient for comparison and analysis.

[0037] Embodiment 2: In the actual tire manufacturing process, the production of sidewall wear-resistant rubber is a continuous dynamic process. In order to accurately determine whether the deformation of the sidewall wear-resistant rubber meets the standard at each moment, the normalized feature matrix is ​​divided into multiple continuous time segments. The length of these time segments can be set according to the actual situation, for example, the length of each time segment is 1 second.

[0038] For each time segment, the dynamic time warping algorithm is used to calculate its similarity with the preset standard deformation mode. The core of the dynamic time warping algorithm is to compare the cumulative distance between the current time segment and the standard deformation mode on the time axis. When calculating the cumulative distance, it is assumed that the current time segment is , the standard mode is , cumulative distance The calculation rule is as follows: ; in, is the Euclidean distance. The Euclidean distance is used to measure the difference between the current time segment and the corresponding point of the standard mode. In actual calculations, for multimodal data such as pressure distribution, temperature gradient and vibration signal, the Euclidean distances of their corresponding points are calculated respectively, and then these distances are comprehensively considered to obtain the total Euclidean distance .

[0039] In the process of calculating the cumulative distance, the algorithm will select the optimal path of the previous moment for extension. Specifically, at each time point, the algorithm will compare the cumulative distances extended from three different directions, select the smallest cumulative distance as the cumulative distance of the current point, and record the extended path. In this way, by continuously extending the optimal path, the cumulative distance of the entire time segment and the standard deformation mode can be obtained.

[0040] When the cumulative distance exceeds the preset deformation tolerance threshold , the time segment is judged to have a significant deviation from the standard mode and is marked as a high-risk interval. The setting of needs to comprehensively consider the precision requirements of the tire manufacturing process and the fluctuations in the actual production process. For example, in the high-precision tire manufacturing process, the deformation tolerance threshold It may be set to a smaller value to ensure the quality of the sidewall wear-resistant rubber; in some production scenarios with relatively low precision requirements, the deformation tolerance threshold In this way, the dynamic time warping algorithm can accurately identify the abnormal deformation of the sidewall wear-resistant rubber during the production process, providing an important basis for subsequent risk assessment and early warning.

[0041] Embodiment 3:

[0042] When constructing the graph neural network model of the sidewall structure, the nodes and edges of the model are first determined. The nodes represent the pressure monitoring points in the wear-resistant rubber area. The distribution of these monitoring points is determined based on the structural characteristics of the sidewall and the actual monitoring needs. For example, more monitoring points are set at key parts of the sidewall where deformation abnormalities are prone to occur to improve the accuracy of monitoring. The edges represent the pressure transmission relationship between adjacent areas, and the propagation of pressure between different areas is simulated by establishing edges.

[0043] During the model update process, the current layer embedding vector of each node is concatenated with the aggregate vectors of its neighboring nodes. The aggregation of neighboring nodes is achieved through a summation operation, that is, the embedding vectors of neighboring nodes are added together to obtain the aggregate vector. For example, for the node , its neighbor nodes are , embedding vectors of these neighbor nodes Sum and get the aggregate vector Then, the node The current layer embedding vector With the aggregation vector Concatenate to get a new vector .

[0044] Next, a linear transformation is performed using a trainable weight matrix. is a parameter in the model, and its value is continuously optimized through training. Perform a linear transformation to obtain Linear transformation can perform weighted summation on vectors and adjust the weights of different features, so that the model can better learn the rules in the data.

[0045] Finally, a nonlinear activation function is applied to generate the node embedding vector of the next layer. Commonly used nonlinear activation functions such as the ReLU function are expressed as By applying a nonlinear activation function, the nonlinear expression ability of the model can be increased, enabling it to learn more complex relationships. Input into the ReLU function to get the node embedding vector of the next layer By repeating the above process, the model can gradually learn the pressure transfer relationship and deformation law between the sidewall wear-resistant rubber areas, thereby accurately outputting the deformation risk probability of each node.

[0046] Embodiment 4: In this embodiment, the reinforcement learning framework adopts the deep Q network (DQN). The deep Q network is an algorithm that combines deep learning and reinforcement learning and can effectively handle high-dimensional state space and complex decision-making problems.

[0047] First, define the state space. The state space is defined as the statistical characteristics of the deformation risk probability, including the mean, variance, maximum, minimum, etc. of the historical deformation risk probability. These statistical characteristics can reflect the overall distribution and change trend of the deformation risk probability. For example, by calculating the mean of the historical deformation risk probability, we can understand the average deformation risk degree of the sidewall wear-resistant rubber over a period of time; by calculating the variance, we can understand the fluctuation of the deformation risk probability.

[0048] Next, define the action space. The action space is the incremental adjustment of the warning threshold, that is, you can choose to increase or decrease the size of the warning threshold. For example, the action space can be set to {-0.1, -0.05, 0, 0.05, 0.1}, which means that the warning threshold can be reduced by 0.1, reduced by 0.05, kept unchanged, increased by 0.05, or increased by 0.1.

[0049] Then, design the reward function. The reward function is designed as .in, is the reward value, is the weight coefficient. and The value of needs to be adjusted according to the actual situation to balance the relationship between the false alarm rate and the number of delay steps. When the value is larger, it means that more attention is paid to reducing the false alarm rate; when When the value is larger, it means that more attention is paid to reducing the number of delay steps. The false alarm rate refers to the proportion of warnings issued when no abnormality actually occurred, and the delay step refers to the time step from the occurrence of an abnormality to the issuance of a warning. By optimizing the reward function, the deep Q network can learn the optimal warning threshold adjustment strategy.

[0050] During the training process, the network parameters are updated through the experience replay mechanism. The experience replay mechanism is an important technology in the deep Q network. It stores the agent's experience (state, action, reward, next state) in an experience replay buffer. During training, a batch of experiences are randomly extracted from the experience replay buffer for learning, which can break the correlation between data and improve the stability and efficiency of learning. For example, the experience replay buffer can store the most recent 1,000 experiences, and 32 experiences are randomly extracted for learning during each training. Through continuous training, the deep Q network can dynamically adjust the warning trigger threshold according to the historical deformation risk probability, generate adaptive warning rules, and improve the performance of the warning system.

[0051] Embodiment 5: In the real-time monitoring and early warning judgment process, the fuzzy logic system plays a core role. Its input variable is the real-time deformation risk probability. and historical average risk . Real-time deformation risk probability The graph neural network risk assessment module outputs in real time, reflecting the deformation risk level of the tire sidewall wear-resistant rubber at the current moment. For example, during the tire production process, if the graph neural network analysis at a certain moment determines the real-time deformation risk probability of a specific area on the tire sidewall, The value is 0.75, indicating that the area is in a state of high deformation risk at the moment. It is obtained by statistically averaging the deformation risk probability over a period of time, which can reflect the long-term risk trend of the sidewall wear-resistant rubber. Assuming that in the past hour, the deformation risk probability of this area fluctuates between 0.5 and 0.7, the historical average risk is calculated. is 0.6.

[0052] The output of the fuzzy logic system is the comprehensive risk index Its fuzzy rule base is carefully constructed based on production experience and data analysis. Rules such as: High and High, then Very high; if Zhongqie Low, then Medium; if Low and risky rate of change High, then High. In actual operation, the fuzzy concept needs to be quantified. , Historical average risk and comprehensive risk index The value range of is divided into three intervals: [0, 0.4) represents "low", [0.4, 0.7) represents "medium", and [0.7, 1] represents "high". For the above example, 0.75 is "high", 0.6 is considered "medium", which does not fully comply with the existing rules, but combined with factors such as the risk change rate, if the risk change rate is at a low level, the comprehensive risk index can be determined based on fuzzy logic reasoning. For "high".

[0053] When the fuzzy logic system calculates the comprehensive risk index After the warning signal is triggered, the control execution unit starts adjusting the pressure parameters of the tire side molding machine. Calculating pressure adjustment .in, is the proportionality coefficient, is the differential coefficient, which needs to be determined according to the actual production situation. In the practice of a tire manufacturer, after many tests and optimizations, it was determined is 0.6, is 0.4. Assuming that the current 0.8, dynamic warning threshold is 0.6, the comprehensive risk index change rate is 0.1, then the pressure adjustment amount Based on this calculation result, the control execution unit sends a command to the sidewall forming machine to increase the corresponding pressure value to correct the abnormal deformation trend of the sidewall wear-resistant rubber and reduce the risk of discounting.

[0054] While adjusting the pressure parameters, the speed parameters also need to be adjusted in coordination. Generally speaking, when the pressure increases, in order to ensure that the sidewall wear-resistant rubber is evenly formed, the speed should be appropriately reduced. In the above example, if the original operating speed of the sidewall forming machine is 12 meters per minute, when the pressure adjustment amount is 0.16, the speed can be adjusted to 10 meters per minute. Through this coordinated regulation of pressure and speed, it can ensure that the sidewall wear-resistant rubber is evenly stressed during the molding process, avoiding rubber surface defects caused by parameter mismatch, such as wrinkles, uneven thickness, etc., fundamentally preventing the sidewall wear-resistant rubber from being folded, and improving the yield rate of tire products.

[0055] In addition, the fuzzy rule base of the fuzzy logic system is not static and needs to be flexibly adjusted according to different tire production processes, equipment characteristics, and raw material differences. For example, when producing radial tires, due to their structural characteristics, the sidewall strength requirements are higher, and the corresponding fuzzy rules may be more stringent and have a lower tolerance for risk. When producing bias tires, the rules can be relatively loose. At the same time, for the proportionality coefficient and the differential coefficient , and also need to be optimized regularly based on production data feedback. If the sidewall wear-resistant rubber is found to be frequently folded over a period of time, it may be that the parameter adjustment is not timely and accurate. At this time, it can be appropriately increased. value to enhance the response speed to risk changes; if there are many false positives, you can adjust value and optimize the pressure adjustment range, so that the entire early warning control system is more in line with actual production needs and continuously ensures the high quality and stability of tire production.

[0056] The present invention also includes an early warning control system for preventing the sidewall wear-resistant rubber from being folded, the system comprising: Data acquisition and preprocessing module: used to collect multimodal data of the sidewall wear-resistant rubber during tire manufacturing, including pressure distribution, temperature gradient and vibration signal, mark the data with timestamps, and perform denoising and feature extraction on the original signal through wavelet transform to obtain a normalized feature matrix; Deformation matching module: divides the normalized feature matrix into continuous time segments, uses the dynamic time warping algorithm to calculate the similarity between each time segment and the preset standard deformation pattern, and generates a deformation matching sequence; Graph neural network risk assessment module: Build a graph neural network model of the sidewall structure, where nodes represent pressure monitoring points in the wear-resistant rubber area, edges represent pressure transmission relationships in adjacent areas, receive deformation matching sequence input, and output deformation risk probability of each node; Warning rule generation module: Based on the reinforcement learning framework, the optimization strategy function is designed, the warning trigger threshold is adjusted according to the historical deformation risk probability, and adaptive warning rules are generated; Real-time monitoring and early warning judgment module: collect multimodal data of the current production cycle in real time, use the data acquisition and preprocessing module and the deformation matching module to obtain the current deformation matching sequence, input the graph neural network risk assessment module to obtain the real-time deformation risk probability distribution, use the fuzzy logic system to process the real-time deformation risk probability distribution, calculate the comprehensive risk index, and determine whether to trigger the early warning signal based on the adaptive early warning rules; Control execution module: if the warning signal is triggered, the pressure parameters and speed parameters of the sidewall forming machine are adjusted, and the adjusted multi-modal data are recorded in the historical database; Model iteration optimization module: Use autoencoders to reconstruct features of the adjusted data in the historical database, calculate the reconstruction error vector, update the standard deformation pattern library, and dynamically correct the weight parameters of the graph neural network based on the Bayesian probability model, combined with the real-time comprehensive risk index and historical reconstruction error.

[0057] The implementation of the system refers to the above embodiment and will not be described in detail in the specification.

[0058] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0059] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A warning control method for preventing the sidewall wear-resistant rubber from being folded, characterized in that: The method is based on dynamic time warping and multimodal data fusion analysis, and includes the following steps: S1. Collect multimodal data of the sidewall wear-resistant rubber during tire manufacturing, including pressure distribution, temperature gradient and vibration signal, mark the data with timestamps, and perform denoising and feature extraction on the original signal through wavelet transform to obtain a normalized feature matrix; S2. Divide the normalized feature matrix into continuous time segments, use the dynamic time warping algorithm to calculate the similarity between each time segment and the preset standard deformation pattern, and generate a deformation matching sequence; S3. Construct a graph neural network model of the sidewall structure, where nodes represent pressure monitoring points in the wear-resistant rubber area and edges represent pressure transmission relationships between adjacent areas; input the deformation matching sequence into the graph neural network and output the deformation risk probability of each node; S4. Based on the reinforcement learning framework, design an optimization strategy function, adjust the warning trigger threshold according to the historical deformation risk probability, and generate adaptive warning rules; S5. Real-time collection of multimodal data of the current production cycle, repeating steps S1 to S2, obtaining the current deformation matching sequence, inputting the graph neural network model of step S3, and obtaining the real-time deformation risk probability distribution; S6. Using a fuzzy logic system to process the real-time deformation risk probability distribution, calculate the comprehensive risk index, and determine whether to trigger a warning signal based on the adaptive warning rule of step S4; S7. If the warning signal is triggered, the pressure parameters and speed parameters of the sidewall molding machine are adjusted by the control execution unit, and the adjusted multimodal data is recorded to the historical database; S8. Reconstruct features of the adjusted data in the historical database using an autoencoder, calculate the reconstruction error vector, and update the standard deformation pattern library; S9. Based on the Bayesian probability model, combined with the real-time comprehensive risk index and historical reconstruction error, the weight parameters of the graph neural network are dynamically corrected to complete the iterative optimization of the model.

2. The method according to claim 1, characterized in that The specific steps of the wavelet transform in step S1 include: S101. performing discrete wavelet decomposition on the pressure distribution signal to extract low-frequency approximate coefficients and high-frequency detail coefficients; S102. Performing a continuous wavelet transform on the temperature gradient signal to generate a time-frequency energy spectrum; S103. Perform wavelet packet decomposition on the vibration signal and select the optimal basis function to construct a frequency band energy distribution matrix; S104. Align the low-frequency approximation coefficients, the time-frequency energy spectrum, and the frequency band energy distribution matrix in time, merge them into a multidimensional feature matrix, and perform Z-score standardization.

3. The method according to claim 1, characterized in that The calculation rule of the dynamic time warping algorithm described in step S2 includes: by comparing the cumulative distance between the current time segment and the standard deformation mode on the time axis, selecting the optimal path of the previous moment for extension, wherein the cumulative distance is calculated by summing up the Euclidean distance between the corresponding points of the current time segment and the standard deformation mode; when the cumulative distance exceeds the preset deformation tolerance threshold, it is determined that there is a significant deviation between the time segment and the standard mode, and it is marked as a high-risk interval; the calculation formula is: ； in, Represents a time segment With standard mode The cumulative distance of is the Euclidean distance; when the cumulative distance exceeds the deformation tolerance threshold , mark the time segment as a high-risk interval.

4. The method according to claim 1, characterized in that: The update rules of the graph neural network model described in step S3 include: concatenating the current layer embedding vector of each node with the aggregation vector of its neighboring nodes, performing a linear transformation through a trainable weight matrix, and then applying a nonlinear activation function to generate the node embedding vector of the next layer, where the aggregation of neighboring nodes is achieved through a summation operation.

5. The method according to claim 1, characterized in that: The reinforcement learning framework described in step S4 adopts a deep Q network (DQN), whose state space is defined as the statistical characteristics of the deformation risk probability, the action space is the incremental adjustment of the warning threshold, and the reward function is designed as: ； in, Reward value, is the weight coefficient, and the network parameters are updated through the experience replay mechanism.

6. The method according to claim 1, characterized in that The input variables of the fuzzy logic system in step S6 include the real-time deformation risk probability and historical average risk , the output is the comprehensive risk index , and its fuzzy rule base is defined as: like High and High, then Very high; like Zhongqie Low, then medium; like Low and If (risk change rate) is high, then high, is the risk change rate.

7. The method according to claim 1, characterized in that The reconstruction error calculation of the autoencoder in step S8 includes: calculating the square error between the input feature matrix and the reconstructed output, and adding the Frobenius norm regularization term of the encoder weight matrix, and optimizing the autoencoder parameters by minimizing the total error.

8. The method according to claim 1, characterized in that The updating process of the Bayesian probability model described in step S9 includes: according to the observation data set consisting of the real-time comprehensive risk index and the historical reconstruction error, combined with the Gaussian prior distribution, the posterior probability distribution of the weight parameter is calculated through the likelihood function to complete the iterative correction of the graph neural network parameters.

9. The method according to claim 1, characterized in that: The parameter adjustment strategy of the control execution unit in step S7 includes: generating a proportional adjustment amount according to the difference between the comprehensive risk index and the dynamic warning threshold, generating a differential adjustment amount in combination with the real-time change rate of the risk index, and outputting the adjustment amount of the pressure parameter after weighted summation. The calculation formula is: ； in, is the pressure adjustment amount, is a real-time comprehensive risk index. are the proportional and differential coefficients, is the dynamic warning threshold, is the rate of change of the comprehensive risk index.

10. An early warning control system for preventing the sidewall wear-resistant rubber from being folded, characterized in that: include: Data acquisition and preprocessing module: used to collect multimodal data of the sidewall wear-resistant rubber during tire manufacturing, including pressure distribution, temperature gradient and vibration signal, mark the data with timestamps, and perform denoising and feature extraction on the original signal through wavelet transform to obtain a normalized feature matrix; Deformation matching module: divides the normalized feature matrix into continuous time segments, uses the dynamic time warping algorithm to calculate the similarity between each time segment and the preset standard deformation pattern, and generates a deformation matching sequence; Graph neural network risk assessment module: Build a graph neural network model of the sidewall structure, where nodes represent pressure monitoring points in the wear-resistant rubber area, edges represent pressure transmission relationships in adjacent areas, receive deformation matching sequence input, and output deformation risk probability of each node; Warning rule generation module: Based on the reinforcement learning framework, the optimization strategy function is designed, the warning trigger threshold is adjusted according to the historical deformation risk probability, and adaptive warning rules are generated; Real-time monitoring and early warning judgment module: collect multimodal data of the current production cycle in real time, use the data acquisition and preprocessing module and the deformation matching module to obtain the current deformation matching sequence, input the graph neural network risk assessment module to obtain the real-time deformation risk probability distribution, use the fuzzy logic system to process the real-time deformation risk probability distribution, calculate the comprehensive risk index, and determine whether to trigger the early warning signal based on the adaptive early warning rules; Control execution module: if the warning signal is triggered, the pressure parameters and speed parameters of the sidewall forming machine are adjusted, and the adjusted multi-modal data are recorded in the historical database; Model iteration optimization module: Use autoencoders to reconstruct features of the adjusted data in the historical database, calculate the reconstruction error vector, update the standard deformation pattern library, and dynamically correct the weight parameters of the graph neural network based on the Bayesian probability model, combined with the real-time comprehensive risk index and historical reconstruction error.

Images