A vulcanized rubber intelligent processing system and method

Abstract

This invention relates to the field of vulcanized rubber processing technology, and discloses an intelligent vulcanized rubber processing system and method. The method acquires a set of physical property parameters of the rubber raw material and real-time operating data of the vulcanization equipment. The set of physical property parameters includes rubber content, filler distribution density, and thermal conductivity, while the real-time operating data includes vulcanization temperature curves, pressure gradient sequences, and mold status indicators. Next, the physical property parameter set and the real-time operating data undergo multimodal feature fusion processing to generate a feature fusion map containing thermodynamic response characteristics and mechanical stress distribution. This map is associated with the phase transition critical point and material deformation trajectory during the vulcanization process. Finally, a process parameter optimization model is invoked to dynamically schedule and analyze the feature fusion map, generating a vulcanization process control instruction set containing a time-series operation sequence based on temperature compensation parameters and pressure adjustment step sizes.

Description

Technical Field

[0001] This invention relates to the field of vulcanized rubber processing technology, specifically to an intelligent vulcanized rubber processing system and method. Background Technology

[0002] In the field of vulcanized rubber processing, the performance of rubber products is closely related to the precision of the vulcanization process, which in turn is affected by both the physical properties of the rubber raw materials and the operating conditions of the equipment. In traditional vulcanization methods, the control of the physical properties of rubber raw materials relies heavily on manual sampling and testing, which can only obtain a few key parameters such as rubber content. For important parameters that affect the uniformity of the vulcanization reaction, such as filler distribution density and thermal conductivity, there is often a lack of real-time and comprehensive monitoring methods. This makes it difficult to detect differences in the physical properties of raw materials in a timely manner, thereby affecting the adaptability of subsequent vulcanization processes.

[0003] In terms of equipment operating condition control, traditional methods often employ fixed vulcanization temperature curves and pressure setpoints, which cannot be adjusted in real time according to the dynamic changes in raw material properties. For example, when the thermal conductivity of rubber raw materials changes due to batch differences, a fixed temperature curve can lead to uneven heating inside the rubber. Some areas may reach the vulcanization critical point prematurely, resulting in over-vulcanization, while other areas may experience incomplete vulcanization due to insufficient temperature, ultimately affecting the mechanical properties of the finished product. Simultaneously, monitoring of mold condition during traditional processing relies primarily on periodic manual inspections, making it impossible to obtain real-time information on mold wear, sealing performance, and other conditions. Even minor mold damage can cause overflow or abnormal deformation of the rubber during vulcanization, reducing product yield, increasing raw material waste, and extending production cycles.

[0004] The optimization of traditional vulcanization process parameters largely relies on the accumulated experience of engineers, adjusting parameters such as temperature and pressure through repeated experiments. This method is not only time-consuming and labor-intensive, but also makes it difficult to achieve precise parameter matching. With the continuous expansion of the application fields of rubber products, the market demands increasingly higher product performance. For example, automobile tires need to have higher wear resistance and aging resistance, and seals need to have more stable elasticity and sealing performance. Traditional processing methods can no longer meet these high-precision and high-stability production requirements. There is an urgent need for a vulcanization rubber processing method that can achieve real-time linkage between raw material properties and equipment operating conditions, and intelligent optimization of process parameters. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent processing method for vulcanized rubber to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent processing method for vulcanized rubber, the method comprising:

[0007] Acquire a set of physical property parameters of rubber raw materials and real-time operating data of vulcanizing equipment. The set of physical property parameters includes rubber content, filler distribution density and thermal conductivity coefficient. The real-time operating data includes vulcanization temperature curve, pressure gradient sequence and mold status indicator.

[0008] The set of physical property parameters and the real-time operating data are subjected to multimodal feature fusion processing to generate a feature fusion map containing thermodynamic response characteristics and mechanical stress distribution. The feature fusion map is associated with the phase transition critical point and material deformation trajectory during the vulcanization process.

[0009] The process parameter optimization model is invoked to dynamically schedule and parse the feature fusion map, generating a vulcanization process control instruction set. The vulcanization process control instruction set includes a timing operation sequence generated based on temperature compensation parameters and pressure adjustment step size.

[0010] Preferably, the step of performing multimodal feature fusion processing on the set of physical property parameters and the real-time operating condition data to generate a feature fusion map containing thermodynamic response features and mechanical stress distribution includes:

[0011] The vulcanization temperature curve and the rubber content are aligned in the time dimension to generate a temperature-component coupling matrix;

[0012] Extract the peak intervals from the pressure gradient sequence and perform spatial mapping processing between the peak intervals and the packing distribution density to generate a pressure-packing correlation tensor.

[0013] The temperature-component coupling matrix and the pressure-filler association tensor are input into a feature cross network and fused using convolution kernels to generate a multi-scale fused feature vector.

[0014] The multi-scale fusion feature vector is subjected to boundary condition constraints based on the mold state identifier to generate the feature fusion map.

[0015] Preferably, the step of calling the process parameter optimization model to dynamically schedule and parse the feature fusion map to generate a vulcanization process control instruction set includes:

[0016] The distribution of phase transition critical points in the feature fusion spectrum is analyzed to generate sulfurization stage division markers and corresponding energy threshold intervals.

[0017] Based on the vulcanization stage division identifier, the time sequence of operations is traversed, and the temperature compensation parameters are calculated by gradient correction.

[0018] The pressure adjustment step size is dynamically weighted according to the mechanical stress distribution to generate a pressure control strategy queue.

[0019] The temperature compensation parameters after gradient correction are integrated with the pressure control strategy queue through instruction encoding to generate the vulcanization process control instruction set.

[0020] Preferably, the method further includes:

[0021] Real-time mold deformation monitoring data is collected during the vulcanization process to generate a process feedback log that includes displacement offset and stress anomaly threshold.

[0022] Extract the abnormal fluctuation features from the process feedback log, and perform pattern matching between the abnormal fluctuation features and the historical vulcanization case library to generate process adjustment instructions;

[0023] The constraint parameters of the process parameter optimization model are updated based on the process adjustment command;

[0024] The updated constraint parameters are injected into the feature cross network, and the energy threshold range in the multi-scale fusion feature vector is recalculated.

[0025] Preferably, the step of extracting abnormal fluctuation features from the process feedback log and performing pattern matching between the abnormal fluctuation features and the historical vulcanization case library to generate process adjustment instructions includes:

[0026] The abnormal fluctuation characteristics are processed by time slicing to generate multiple fluctuation event segments and their corresponding temperature and pressure joint curves;

[0027] The pre-trained anomaly attribution model is invoked to perform root cause analysis on each fluctuation event segment, generating a set of classification labels including uneven filler distribution, sudden temperature change, and mold aging.

[0028] Retrieve process correction templates that match the classification tag set from the historical vulcanization case library to generate a candidate adjustment strategy set;

[0029] Based on the similarity ranking between the temperature-pressure joint curve and the candidate adjustment strategy set, the strategy with the highest matching degree is selected to generate the process adjustment instruction.

[0030] Preferably, the step of calling the pre-trained anomaly attribution model to perform root cause analysis on each fluctuation event segment generates a set of classification labels including uneven filler distribution, sudden temperature change, and mold aging, including:

[0031] The fluctuation event segments are processed by keyframe extraction to generate a stress mutation frame sequence and timestamp index;

[0032] Calculate the difference in thermal expansion coefficients between adjacent frames in the stress abrupt change frame sequence to generate a material deformation acceleration vector;

[0033] The material deformation acceleration vector is input into the multipath convolutional subnet of the anomaly attribution model for local feature extraction, generating a spatial anomaly response map.

[0034] The timestamp index is subjected to window sliding processing to generate a histogram of time-dimensional anomaly distribution;

[0035] The spatial anomaly response map and the temporal dimension anomaly distribution histogram are subjected to feature cross-validation to generate a root cause probability distribution matrix.

[0036] Extract the peak indices of the probability of uneven packing distribution, the probability of sudden temperature change, and the probability of mold aging from the root cause probability distribution matrix.

[0037] The preset probability threshold is dynamically compared based on the peak index to generate a set of classification labels with weighted coefficients.

[0038] Preferably, the method further includes:

[0039] Virtual operating condition disturbance parameters are injected before the vulcanization process is started. These virtual operating condition disturbance parameters are used to simulate pressure pulsation and temperature drift scenarios.

[0040] Monitor the stability processing results of the process parameter optimization model on the disturbed operating conditions, and generate fault tolerance evaluation indicators;

[0041] When the fault tolerance evaluation index is lower than the preset threshold, the parameter adaptive update mechanism of the feature cross network is triggered;

[0042] The convolutional kernel weights of the feature cross network are iteratively optimized based on the difference in physical property parameters before and after the perturbation.

[0043] Preferably, the iterative optimization of the convolutional kernel weights of the feature cross network based on the difference in physical property parameters before and after the perturbation includes:

[0044] Tensor alignment is performed on the difference data of physical property parameters before and after the disturbance to generate a physical property offset matrix;

[0045] Extract the rate of change of thermal conductivity and the packing distribution offset vector from the physical property offset matrix;

[0046] The rate of change of thermal conductivity coefficient and the packing distribution offset vector are concatenated to generate the physical property loss function input tensor.

[0047] The prediction error gradient of the feature cross network is calculated based on the input tensor of the physical property loss function.

[0048] The prediction error gradient is smoothed by momentum to generate the convolution kernel weight update direction vector;

[0049] The convolution kernel parameters in the feature cross network are adjusted based on the convolution kernel weight update direction vector.

[0050] Preferably, the method further includes:

[0051] Build a cross-equipment process adapter and analyze the differences in control protocols among different vulcanization equipment;

[0052] The vulcanization process control instruction set is converted into a sequence of low-level drive instructions executable by the target device;

[0053] The phase synchronization relationship and exception handling logic of the timing operation sequence are preserved during the conversion process;

[0054] Inject real-time response parameters that match the target device to generate device-independent process control bytecode.

[0055] Preferably, the present invention also includes an intelligent processing system for vulcanized rubber, comprising a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the steps of the above method.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] This intelligent vulcanized rubber processing method achieves comprehensive perception of key influencing factors during vulcanization by acquiring a set of physical property parameters of rubber raw materials and real-time operating data of vulcanization equipment. Compared with traditional processing methods that can only monitor a few raw material parameters and fixed equipment operating conditions, this method covers physical property parameters such as rubber content, filler distribution density, and thermal conductivity, as well as operating condition data such as vulcanization temperature curves, pressure gradient sequences, and mold status indicators. It can more comprehensively reflect the characteristics of raw materials and the operating status of equipment, providing richer and more accurate basic information for subsequent process adjustments and avoiding the problem of insufficient process adaptability caused by the lack of key parameters.

[0058] In the multimodal feature fusion processing stage, this method fuses physical property parameters with operating condition data to generate a feature fusion map that includes thermodynamic response characteristics and mechanical stress distribution, and correlates the phase transition critical point and material deformation trajectory during vulcanization. This process breaks the limitation of the independence between raw material parameters and equipment operating conditions in traditional processing, and can clearly show the impact of changes in raw material properties on equipment operating conditions, as well as the effect of adjusting equipment operating conditions on the rubber vulcanization process. For example, when the filler distribution density of the raw material is high, the feature fusion map can intuitively observe the changes in the internal heat conduction rate of the rubber, and the resulting shift in the phase transition critical point, thus providing a clear direction for subsequent process parameter adjustments and avoiding blind adjustments caused by unclear parameter correlations in traditional processing.

[0059] By calling a process parameter optimization model to dynamically schedule and analyze the feature fusion map, a time-series operation sequence containing temperature compensation parameters and pressure adjustment step sizes is generated, achieving intelligent and dynamic optimization of vulcanization process parameters. Traditional processing methods rely on fixed parameters or manual experience for adjustment, which cannot respond promptly to dynamic changes in raw material properties and equipment conditions. This method, however, uses model analysis to automatically generate control commands adapted to the current state based on real-time information in the feature fusion map. For example, when a decrease in the thermal conductivity of the rubber raw material is detected, the model automatically calculates the temperature compensation parameter and adjusts the vulcanization temperature curve to ensure uniform heating inside the rubber. When the mold status indicator shows a slight decrease in mold sealing performance, the model generates a corresponding pressure adjustment step size and appropriately adjusts the pressure gradient sequence to avoid rubber overflow or abnormal deformation. This dynamic optimization method ensures that process parameters always maintain optimal matching with current raw material properties and equipment conditions, effectively reducing the occurrence of over-vulcanization and under-vulcanization, and improving the stability of product performance.

[0060] This intelligent processing method can also significantly improve production efficiency and reduce resource consumption. In traditional processing, product defects caused by untimely or inaccurate parameter adjustments require rework or scrapping, wasting raw materials and extending the production cycle. This method, through precise parameter optimization, can improve the first-pass yield and reduce raw material consumption. Real-time monitoring and dynamic adjustment of operating conditions also avoid downtime for maintenance due to undetected equipment malfunctions, shortening the production cycle. Furthermore, this method eliminates the need for repeated trials and adjustments based on engineers' experience, reducing reliance on human experience and minimizing human error. This makes the production process more standardized and repeatable, facilitating large-scale, standardized production and better meeting the diverse performance requirements of rubber products in different application scenarios. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent processing method for vulcanized rubber described in this invention.

[0062] Figure 2 A flowchart for multimodal feature fusion processing;

[0063] Figure 3 The flowchart shows the feedback and model update process during the vulcanization execution.

[0064] Figure 4 The flowchart for virtual operating condition disturbance and model fault tolerance optimization. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Please see Figure 1 This invention provides an intelligent processing system and method for vulcanized rubber, the method comprising:

[0067] The process involves acquiring a set of physical properties of the rubber raw material, including rubber content, filler density, and thermal conductivity. Simultaneously, real-time operating data of the vulcanization equipment is collected, including vulcanization temperature curves, pressure gradient sequences, and mold status indicators. Subsequently, multimodal feature fusion processing is performed on the physical property parameter set and the real-time operating data to generate a feature fusion map that simultaneously incorporates thermodynamic response characteristics and mechanical stress distribution. This map correlates the phase transition critical point and material deformation trajectory during the vulcanization process. Finally, a process parameter optimization model is invoked to dynamically schedule and analyze the feature fusion map, generating a vulcanization process control instruction set. This instruction set contains a time-series operation sequence based on temperature compensation parameters and pressure adjustment step sizes, directly driving the vulcanization equipment to perform precise process control.

[0068] Example 1: See Figure 2 After acquiring physical properties such as rubber content, filler density, and thermal conductivity, as well as real-time operating data such as vulcanization temperature curves, pressure gradient sequences, and mold status indicators, the system initiates multimodal feature fusion processing. First, a time-dimensional alignment operation is performed, matching each timestamp data point of the vulcanization temperature curve with its corresponding rubber content parameter. This alignment process relies on a high-precision time synchronization algorithm, which compensates for any minor time delays that may occur during data acquisition, ensuring that each temperature reading is accurately correlated with the material's compositional characteristics at that moment. Through this alignment, the system constructs a temperature-composition coupling matrix, which not only records the temperature change over time but also reflects the intrinsic relationship between the material composition and its temperature response.

[0069] The processing of the pressure gradient sequence focuses on extracting spatial characteristics. The system uses a peak detection algorithm to scan the entire pressure sequence, identify all extreme points of pressure changes, and delineate peak intervals centered on these points. These peak intervals represent key stages of mechanical stress during vulcanization. Subsequently, the system spatially maps these pressure characteristics to the filler distribution density data. The mapping process considers the physical structure of the mold and the actual distribution of the filler. Through a three-dimensional spatial interpolation algorithm, a quantitative correlation is established between the pressure data and the filler distribution data in the spatial dimension, ultimately generating a pressure-filler correlation tensor that reflects the interaction between pressure and filler.

[0070] The Feature Cross-Network receives the generated temperature-component coupling matrix and pressure-filler correlation tensor as input. This network employs a multi-channel convolutional architecture, with each channel specifically processing feature information from different modalities. Internally, data from different sources are fused using convolutional kernels with shared weights. This design enables the network to capture the nonlinear interactions between multiple factors such as temperature, pressure, and component distribution. Convolutional operations are performed at multiple scales, progressively extracting and integrating features from local subtleties to global macroscopic features, ultimately outputting a fused vector containing rich feature information.

[0071] The mold status identifier acts as a boundary constraint in this process. Based on the mold's current state parameters, such as wear coefficient and mold closing accuracy, the system verifies the validity and limits the range of the fused feature vector. By applying a constraint function based on physical rules, the system ensures that all feature values ​​are within a reasonable physical range, thereby generating the final feature fusion map. This map, in a structured data format, fully describes the thermodynamic response characteristics and mechanical stress distribution during the vulcanization process, and clearly identifies key information such as the phase transition critical point and material deformation trajectory.

[0072] When analyzing the feature fusion spectrum, the process parameter optimization model first focuses on the distribution pattern of phase transition critical points. A density clustering algorithm is used to identify clusters of phase transition points in the spectrum, and based on the distribution characteristics of these regions, the entire vulcanization process is divided into several stages with different characteristics. Based on the stage division results, the system traverses the time-series operation sequence and performs gradient correction on the temperature compensation parameters. The correction process employs a gradient-based optimization algorithm to calculate the difference between the current temperature setting and the ideal temperature curve, and adjusts the compensation parameters accordingly. This dynamic adjustment ensures that temperature control can adapt to the specific needs of different vulcanization stages.

[0073] Mechanical stress distribution data is used for dynamic weighted processing of pressure regulation step size. The system analyzes the uniformity and intensity characteristics of stress distribution and adjusts the precision of pressure control in real time through a weighting function. In areas with uneven stress distribution or concentration, the system automatically reduces the step size to achieve finer pressure control; while in areas with uniform distribution, a larger step size is used to improve control efficiency. The pressure control strategy queue generated in this way enables precise adaptive control of vulcanization pressure. The system integrates the gradient-corrected temperature compensation parameters with the dynamically weighted pressure control strategy through instruction encoding. The encoding process uses a standardized industrial control protocol to convert all process parameters into a sequence of instructions that the equipment can directly execute. These instructions are strictly arranged in chronological order and maintain the synchronization relationship between parameters, ultimately forming a complete vulcanization process control instruction set. This instruction set is transmitted to the control system of the vulcanization equipment via an industrial bus, driving the equipment to perform precise vulcanization process operations.

[0074] Example 2: See Figure 3 During the vulcanization process, the system continuously collects mold deformation data through a distributed sensor network. Laser displacement sensors installed at key locations on the mold monitor changes in the mold closing gap with an accuracy of 0.01 mm, while an embedded stress sensor array captures the pressure distribution inside the mold in real time. These sensors generate raw deformation data streams at a sampling frequency of 100 Hz. After noise is filtered by the data preprocessing module, a structured process feedback log is generated. This log includes quantified values ​​of displacement offset, stress anomaly threshold markers, and corresponding timestamps, where the anomaly threshold is dynamically set based on the material's yield strength. For example, when the displacement offset in a certain area exceeds 0.15 mm for three consecutive sampling cycles, the system automatically marks the stress anomaly threshold for that time period.

[0075] The extraction of abnormal fluctuation features employs an adaptive sliding window mechanism. The system uses the timestamps in the process feedback log as a baseline, and automatically extends the analysis window forward by 200 milliseconds and backward by 500 milliseconds when a stress anomaly threshold is detected. Displacement and stress data within each window are decomposed into components of different frequency bands using wavelet transform. High-frequency components exceeding three times the standard deviation of the baseline fluctuation range are extracted as abnormal fluctuation features. These features, along with the corresponding temperature-pressure joint curves, are encapsulated into independent fluctuation event segments. Each segment includes a time start marker, duration, and feature vector.

[0076] The pre-trained anomaly attribution model employs a three-channel input architecture to process fluctuating event segments. The first channel receives spatial distribution data from displacement sensors and extracts deformation region features through a 3D convolutional layer. The second channel analyzes the time-frequency characteristics of stress sensors and uses a long short-term memory network to capture temporal dependencies. The third channel processes the temperature-pressure joint curve and uses a fully connected layer to identify thermo-coupling patterns. The model's output layer contains three parallel softmax classifiers, corresponding to probability predictions for three root causes: uneven filler distribution, sudden temperature changes, and mold aging. When a fluctuating event segment is input, the model may output a probability distribution of [0.85, 0.10, 0.05], indicating that the anomaly has an 85% probability of originating from a filler distribution problem.

[0077] The historical vulcanization case database is stored in a graph database. Each case contains four dimensions of data: anomaly feature vector, process parameters, treatment measures, and effect evaluation. When the system matches the feature vector of the current fluctuation event segment with the case database, it uses an improved Euclidean distance algorithm: assigning a weight of 0.6 to temperature-related features, 0.3 to pressure features, and 0.1 to displacement features. The matching process returns the top 5 historical cases with a similarity greater than 0.8, and the process correction measures in these cases are extracted as a candidate adjustment strategy set.

[0078] In the strategy selection phase, a multi-dimensional evaluation matrix is ​​established. The system dynamically aligns the current temperature-pressure joint curve with the execution curves of each candidate strategy through time warping, calculating the curve shape similarity. Simultaneously, it compares the differences between the current equipment operating conditions and historical cases. Finally, it evaluates the strategy implementation complexity. Through a three-factor weighted scoring system, the strategy with the highest overall score is selected to generate process adjustment instructions. For example, an instruction might require "increasing the holding pressure by 5% in region C while simultaneously reducing the heating rate by 0.5℃ / s."

[0079] The constraint updates for the process parameter optimization model are implemented through a parameter mapping interface. The system parses the operational parameters in the process adjustment commands and converts them into model constraint parameters. For example, pressure adjustments are converted into boundary values ​​for stress distribution constraints, and temperature change rates are converted into coefficient correction terms for the heat conduction equation. These updated constraint parameters are injected into the model solver, altering the shape of the objective function.

[0080] After receiving the updated constraint parameters, the feature cross-network initiates the feature recalculation process. The network first freezes the convolutional kernel weights and adjusts only the normalization parameters of the feature fusion layer. For the energy threshold interval in the multi-scale fused feature vector, the system employs an interval iterative shrinking algorithm: centering on the current threshold, the search space is expanded bidirectionally with a step size of 0.5%. The feature matching degree under each candidate threshold is calculated through forward propagation, and the point with the highest matching degree is selected as the new interval boundary. The entire process is completed within 50 milliseconds, ensuring continuous real-time control.

[0081] When the system detected a 0.18 mm displacement at the lower right corner of the mold, it automatically triggered the anomaly handling process. Analysis identified a filler distribution probability of 0.78 in this area, and a review of historical cases revealed that similar issues had been resolved by adjusting the holding pressure curve. The system generated an adjustment command to "increase the holding pressure in stage three by 8% and extend the holding time by 3 seconds." After model updates, the recalculated feature fusion spectrum showed a 23% improvement in stress distribution uniformity in this area. The entire dynamic adjustment process was completed within the vulcanization cycle without interrupting the production process.

[0082] Through a collaborative mechanism of multi-source sensor data fusion, intelligent root cause analysis, case-driven decision-making, and online model updates, the vulcanization process has achieved autonomous optimization capabilities. Each technical aspect adopts a modular design; for example, the anomaly attribution model supports incremental learning, and the historical case library has an automatic cleaning mechanism to ensure the system's continuous evolutionary characteristics. This design philosophy enables vulcanization process control to move from static preset to dynamic response, adapting to complex and ever-changing industrial production environments.

[0083] Example 3: After receiving a fragment of a fluctuation event, the system first performs keyframe extraction processing. This processing uses a gradient-based edge detection algorithm to scan and analyze the stress data stream in the fragment. The algorithm sets a stress change rate threshold ζ = 15 MPa / s. When the stress change rate of two consecutive sampling points exceeds this threshold, it is marked as the stress mutation start frame. Tracking backward from the start frame, the frame is marked as the end frame when the change rate is less than 5% of the threshold ζ, thus generating a stress mutation frame sequence containing start and end timestamps. Each mutation frame is accompanied by a timestamp index accurate to milliseconds, forming a time-stress intensity correspondence table.

[0084] In the thermal expansion coefficient difference calculation stage, the system performs frame-by-frame analysis on the stress abrupt change frame sequence. For two adjacent frames... and The difference in their coefficients of thermal expansion The calculation method is as follows:

[0085]

[0086] in: Indicates the time interval The change in internal thermal strain. This represents the temperature change within the corresponding time interval. The calculation process is based on the thermodynamic constitutive relations of the material, solving for the instantaneous thermal expansion characteristics using real-time acquired temperature-strain coupled data. The system performs this calculation on all adjacent frames in the sequence, generating an acceleration vector describing the change in the material's deformation rate. The magnitude of this vector reflects the severity of deformation, and its direction indicates the deformation trend.

[0087] The material deformation acceleration vector is input into a multi-path convolutional subnetwork of the anomaly attribution model. This subnetwork employs a three-branch parallel architecture: the first branch uses a 1×1 convolutional kernel to extract vector magnitude features; the second branch uses a 3×3 convolutional kernel to capture local change patterns; and the third branch uses a 5×5 convolutional kernel to obtain global context information. The outputs of each branch are batch normalized and then weighted and concatenated in a feature fusion layer. The fused feature tensor is then enhanced by a spatial attention module to improve the response in important regions, ultimately generating a spatial anomaly response map. This map is displayed as a heatmap, indicating the spatial distribution intensity of anomalous events, with red areas representing highly sensitive regions where the anomaly response value exceeds 0.8.

[0088] The time dimension analysis employs a sliding window processing mechanism. The system defines a time-dimensional analysis based on the timestamp index of stress mutation frames. The center timestamp of the adaptive time window is the time point when the abnormal fluctuation event is first detected by the system, which is used to anchor the time range benchmark for anomaly analysis. The time window is set to its half-width value, dynamically adjusted based on the actual duration of the abnormal event, thus creating an adaptive time window. ,in The algorithm dynamically adjusts based on the duration of the anomaly, ranging from 50 to 200 milliseconds. Within each window, the frequency, duration, and intensity integral values ​​of the anomaly response are statistically analyzed to generate a time-dimensional anomaly distribution histogram. This histogram contains 24 time bins, each representing the cumulative anomaly energy value within a 5-millisecond time interval.

[0089] The feature cross-validation step establishes a correlation matrix between spatial and temporal features, resamples the spatial anomaly response map to the same temporal resolution as the temporal histogram, and then calculates the mutual information value between the two at the same time:

[0090]

[0091] in: Represents a sequence of spatial anomaly response values. Represents a sequence of outliers over time. For joint probability distribution, and This represents the marginal probability distribution. The correlation strength of spatial-temporal features is obtained through this calculation, serving as the weighting basis for subsequent root cause determination.

[0092] The generation of the root cause probability distribution matrix employs a Bayesian inference framework. The system pre-establishes conditional probability models for three types of root causes: uneven packing distribution corresponds to a strong spatial response with a long duration; sudden temperature changes exhibit a sharp temporal response with a wide spatial distribution; and mold aging displays a spatially localized but periodically occurring characteristic. The extracted spatial anomaly response map and temporal anomaly distribution histogram are input into this framework, and the probability of occurrence for each type of root cause is calculated using maximum a posteriori probability estimation, forming a 3×N probability distribution matrix (where N is the number of anomaly events).

[0093] Peak index extraction employs a multi-scale extremum detection algorithm. For each row of the probability distribution matrix (corresponding to a root cause), a Gaussian filter is first used for smoothing, and then a sliding window is used to find local maxima. The window size is adaptively adjusted according to the characteristics of the probability distribution: a larger window (7 data points wide) is used for regions with gentle probability gradient changes, and a smaller window (3 data points wide) is used for regions with steep gradients. The coordinates of each extremum point are recorded as the peak index, along with the probability value of that point and the probability gradient of its neighboring points.

[0094] The dynamic comparison processing introduces an adaptive threshold mechanism. The preset probability threshold is initially set to 0.6, but it is dynamically adjusted based on the overall characteristics of the abnormal event: when the duration of the abnormality exceeds 300 milliseconds, the threshold is lowered to 0.55; when the spatial distribution range exceeds 30% of the mold area, the threshold is raised to 0.65. The system compares the probability value corresponding to each peak index with the current threshold. Peaks exceeding the threshold are retained and assigned a weight coefficient, which is calculated linearly based on the magnitude of the probability exceeding the threshold.

[0095] The final generated set of classification labels uses a multi-dimensional vector representation. Each label contains four elements: root cause type (uneven filler distribution, sudden temperature change, or mold aging), confidence probability (a value between 0 and 1), weight coefficient (an adjustment factor between 0.5 and 1.5), and scope of influence (a three-level classification of local / regional / global). For example, the output might be the label combination "uneven filler distribution: 0.82: 1.2: region," indicating that the anomaly has an 82% probability of being caused by a filler distribution problem, requiring adjustment with a weight coefficient of 1.2, and the scope of influence is a certain region of the mold.

[0096] When the system detects abnormal stress in the tread pattern area, it extracts and locks the anomaly at the 128th second of vulcanization using keyframe extraction. The calculated difference in the coefficient of thermal expansion during this period is 0.15 × 10⁻⁻⁻⁻⁶. 6 / K, significantly higher than the normal value of 0.05×10⁻ 6 / K, where K is the thermodynamic temperature unit and the basic unit of temperature in the International System of Units (SI). 1 K equals 1 °C in temperature change and is used to quantify the impact of temperature changes on the thermal expansion properties of materials. The spatial anomaly response map shows a response intensity of 0.9 in the grooved pattern region, and the time histogram shows the anomaly lasts for 180 milliseconds. Feature cross-validation yields a mutual information value of 0.76, indicating a high correlation between spatiotemporal features. Root cause probability calculations show a probability of 0.88 for uneven filler distribution, 0.07 for sudden temperature changes, and 0.05 for mold aging, ultimately generating a high-confidence label for the filler distribution anomaly.

[0097] Example 4: See Figure 4 The virtual operating condition disturbance parameters are generated using a multi-mode hybrid strategy. The system maintains a database containing 12 typical disturbance modes, each with corresponding parameter disturbance rules. Pressure pulsation simulation employs random amplitude modulation, with the base pressure value... Superimposed random fluctuations ,in It follows a mean of 0 and a standard deviation of 0.15. The normal distribution. Temperature drift simulation introduces slowly varying disturbance components within a set temperature curve. Superimposed Periodic perturbation, amplitude The disturbance parameters are randomly generated within the temperature range of 2-8℃, with frequencies ranging from 0.001-0.01Hz (low frequency range). For each training iteration, 3-5 disturbance patterns are randomly selected and injected. The table below shows an example of the disturbance parameters generated during a particular training iteration.

[0098] Table 1: Virtual Operating Condition Disturbance Parameter Configuration

[0099] Disturbance type Parameter range Time mode High-frequency pressure pulsation Amplitude ±12% of rated pressure Random burst, lasting 3-8 seconds Temperature step drift Offset +5℃ / -3℃ Step change every 120 seconds thermal conductivity attenuation Attenuation rate 0.8-0.95 times Continuous effect throughout the process Packing distribution offset Regional unevenness increased by 30%. Phase 2 will take effect.

[0100] The fault tolerance assessment index is calculated using a three-dimensional metric. The stability dimension records the fluctuation coefficient of the output commands from the process parameter optimization model and calculates the standard deviation change rate of the vulcanization temperature control commands before and after disturbances. The consistency dimension compares the structural similarity index of the feature fusion map under standard and disturbed operating conditions. The reliability dimension counts the frequency with which the model output commands exceed the equipment safety boundary. These three indicators are assigned weight coefficients of 0.4, 0.3, and 0.3 respectively, and a comprehensive calculation yields a fault tolerance assessment value in the 0-1 range. The system sets a fault tolerance threshold of 0.85; when the assessment value falls below this threshold, a network update process is automatically triggered.

[0101] The adaptive parameter update of the feature cross network comprises seven operational stages. First, tensor alignment of the physical property parameter difference data is performed, resampling the glue content distribution map, filler density field, and thermal conductivity matrix before and after perturbation to unify them to the same spatial resolution and time scale. The aligned data is then used to generate a physical property offset matrix through matrix subtraction, where each element records the parameter change at a specific location before and after perturbation. The offset feature extraction stage focuses on key physical properties; the rate of change of thermal conductivity is obtained by calculating the mean change in the heat conduction channel region within the offset matrix. The filler distribution offset vector is extracted using principal component analysis, extracting the changes in the first three principal component directions from the filler distribution difference data. These two feature vectors capture the essential changes in the material's thermal response and structural properties, respectively.

[0102] The input tensor of the physical property loss function is constructed through a feature concatenation layer, which concatenates the vector of the rate of change of thermal conductivity coefficient with the offset vector of filler distribution along the channel dimension to form a dual-channel feature map. Each channel's data undergoes min-max normalization to eliminate the influence of dimensional differences. The concatenated tensor contains both spatial and feature dimension information, fully describing the physical property change pattern caused by perturbations. The prediction error gradient is calculated using a backpropagation mechanism. The feature cross-network performs forward propagation on the concatenated tensor, outputting the predicted physical property change. The mean square error is calculated between this predicted value and the actual offset matrix, and the error signal propagates backward along the network. At the convolutional layer nodes, the system records the gradient contribution value of each convolutional kernel, forming an initial gradient map. Momentum smoothing optimizes the gradient update direction. The system maintains a gradient momentum buffer, storing the gradient history data from the last five iterations. The new prediction error gradient is weighted and averaged with the historical gradients, with the current gradient assigned a weight of 0.7 and the historical gradients assigned a weight of 0.3. The smoothed gradient effectively suppresses random fluctuations, generating a more stable convolutional kernel weight update direction vector. The convolutional kernel parameters are adjusted using a phased update strategy. The 64 convolutional kernels in the network are divided into three groups based on their update direction vectors: high-frequency feature extraction kernels use a larger learning rate (0.01) to quickly respond to changes in physical properties; spatial correlation kernels use a medium learning rate (0.005); and background feature kernels maintain a smaller learning rate (0.001). The update process uses a batch iterative approach, updating 20% ​​of the convolutional kernel parameters each time, and completing all parameter adjustments in five iterations.

[0103] In an anti-interference training exercise targeting large rubber gaskets, the system was injected with a combination of disturbances: pressure fluctuations of ±10%, temperature drift of +6℃, and a 25% increase in filler distribution unevenness. The initial fault tolerance assessment value was only 0.72, triggering the network update mechanism. Property shift analysis showed a maximum decrease of 18% in thermal conductivity, with filler distribution shift concentrated in region B. After three rounds of iterative updates, the fault tolerance assessment value under the same disturbances increased to 0.91. The updated network successfully maintained the stability of vulcanization quality in real production when pressure sensor failure occurred. By simulating parameter drift under extreme operating conditions, the system proactively established a model library to handle abnormal situations. The parameter update mechanism of the characteristic cross-network enabled the model to adapt online, effectively reducing the impact of real-world factors such as equipment aging and raw material fluctuations on process stability.

[0104] Example 5: The cross-device process adapter includes a protocol parsing engine and an instruction conversion module. The protocol parsing engine has a built-in device feature library that stores the communication protocol characteristics of various vulcanizing equipment. When connecting new equipment, the engine automatically scans the equipment's control instruction set, identifying the instruction format, data encoding method, and verification rules. For a certain brand of flat vulcanizing machine, parsing reveals that it uses the Modbus-RTU protocol, with temperature control instructions in 16-bit integer format and pressure control in 32-bit floating-point format. The engine extracts these protocol elements to generate a device protocol description file, which marks key feature points: the instruction start character is 0x5A, data bits use big-endian order, and the verification method is CRC-16.

[0105] After receiving the vulcanization process control instruction set, the instruction conversion module initiates a multi-stage conversion process. Internally, the module maintains an instruction mapping table, which establishes the correspondence between high-level instructions and low-level operations. For example, the "stage heating rate control" instruction is mapped to the equipment-specific "temperature ramp setting" function, whose parameters include the target temperature, heating time, and overshoot limit value. The conversion process first decomposes the timing operation sequence, matching each operation unit with the mapping table to generate preliminary drive instruction fragments.

[0106] The phase synchronization relationship is maintained using timestamp anchoring technology. The system marks key time nodes in the original instruction set, such as the pressure regulation start point and the temperature holding stage transition point. During the transition, synchronization instructions are inserted at these nodes to create a time base reference between devices. For a multi-station vulcanization production line, the system marks the synchronization relationship between "station 2 pressurization" and "station 3 depressurization" in the time sequence. After the transition, it generates timestamped instructions: station 2 executes "Pressure_Set(15.8MPa,T+0ms)", and station 3 executes "Pressure_Release(T+200ms)", ensuring that the 200-millisecond phase difference is accurately achieved.

[0107] The migration of exception handling logic adopts a rule inheritance mechanism. The exception handling in the original instruction set includes a three-level response for temperature over-limit: the primary response is to reduce heating power, the secondary response is to activate the cooling system, and the emergency response is to execute emergency mold opening. The conversion module analyzes the exception response capability of the target equipment and maps each level of response to specific operations. When the target equipment does not have an independent cooling system, the secondary response is automatically converted into a combined instruction of "cutting off the heating power + starting the fan". The logic migration process maintains the exception handling hierarchy unchanged, only replacing the underlying execution method. The injection of real-time response parameters addresses the differences in equipment characteristics. The system establishes a response parameter file for each type of equipment, including dynamic parameters such as instruction transmission delay, actuator response time, and sensor sampling period. When a pressure adjustment instruction is sent to a hydraulic vulcanizing machine, the system detects a 300-millisecond delay in the hydraulic system of the equipment and automatically adds time compensation to the instruction sequence: correcting the theoretical execution time T to T-300ms. At the same time, according to the sampling period of the equipment's pressure sensor (50 milliseconds), the pressure feedback detection interval is adjusted to match the actual response characteristics of the equipment.

[0108] The generation of process control bytecode employs intermediate representation technology. The converted instruction sequence is compiled into device-independent bytecode, with each bytecode consisting of four parts: opcode, timestamp, parameter field, and checksum. The opcode uses a unified encoding system, such as 0x01 representing temperature control and 0x02 representing pressure control; the timestamp records the relative execution time; the parameter field stores the normalized control parameters; and the checksum is generated using an adaptive algorithm. The bytecode is interpreted and executed by the device driver layer, which translates the bytecode into native instructions based on the target device type. In one conversion example, the original instruction "temperature compensation +3.5℃" was compiled into bytecode [0x01,T+1200ms,0.035,0x8A2F]. On an electric vulcanizing machine, this is interpreted as "SET_TEMP_OFFSET+3.5", while on a steam-type machine, it is interpreted as "STEAM_VALVE_OPEN+5%".

[0109] This implementation method has been validated for effectiveness in the retrofitting of a rubber conveyor belt production line, which includes three different generations of vulcanizing equipment: a hydraulic flat vulcanizing machine (2005), an electric heating automatic vulcanizing line (2015), and an electromagnetic induction vulcanizing unit (2020). The process adapter successfully converted unified process instructions into dedicated instruction sequences for the three types of equipment, maintaining a vulcanizing pressure curve synchronization error of less than 0.5 MPa and a temperature trajectory deviation of no more than 2°C. The bytecode transmission mechanism enables process adjustment instructions to be synchronously sent to all equipment within 50 milliseconds, achieving coordinated control of the entire line. The protocol parsing layer encapsulates equipment differences, the instruction conversion layer maintains the integrity of the process logic, and the bytecode execution layer implements a unified interface. Through a triple mechanism of time anchoring, rule inheritance, and parameter compensation, it is ensured that process intentions are executed equivalently on different equipment.

[0110] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent processing method for vulcanized rubber, characterized in that, include: Acquire a set of physical property parameters of rubber raw materials and real-time operating data of vulcanizing equipment. The set of physical property parameters includes rubber content, filler distribution density and thermal conductivity coefficient. The real-time operating data includes vulcanization temperature curve, pressure gradient sequence and mold status indicator. The set of physical property parameters and the real-time operating data are subjected to multimodal feature fusion processing to generate a feature fusion map containing thermodynamic response characteristics and mechanical stress distribution. The feature fusion map is associated with the phase transition critical point and material deformation trajectory during the vulcanization process. The process parameter optimization model is invoked to dynamically schedule and parse the feature fusion map, generating a vulcanization process control instruction set. The vulcanization process control instruction set includes a timing operation sequence generated based on temperature compensation parameters and pressure adjustment step size.

2. The intelligent processing method for vulcanized rubber as described in claim 1, characterized in that, The step of performing multimodal feature fusion processing on the set of physical property parameters and the real-time operating condition data to generate a feature fusion map containing thermodynamic response features and mechanical stress distribution includes: The vulcanization temperature curve and the rubber content are aligned in the time dimension to generate a temperature-component coupling matrix; Extract the peak intervals from the pressure gradient sequence and perform spatial mapping processing between the peak intervals and the packing distribution density to generate a pressure-packing correlation tensor. The temperature-component coupling matrix and the pressure-filler association tensor are input into a feature cross network and fused using convolution kernels to generate a multi-scale fused feature vector. The multi-scale fusion feature vector is subjected to boundary condition constraints based on the mold state identifier to generate the feature fusion map.

3. The intelligent processing method for vulcanized rubber as described in claim 2, characterized in that, The process parameter optimization model is invoked to dynamically schedule and parse the feature fusion map, generating a vulcanization process control instruction set, including: The distribution of phase transition critical points in the feature fusion spectrum is analyzed to generate sulfurization stage division markers and corresponding energy threshold intervals. Based on the vulcanization stage division identifier, the time sequence of operations is traversed, and the temperature compensation parameters are calculated by gradient correction. The pressure adjustment step size is dynamically weighted according to the mechanical stress distribution to generate a pressure control strategy queue. The temperature compensation parameters after gradient correction are integrated with the pressure control strategy queue through instruction encoding to generate the vulcanization process control instruction set.

4. The intelligent processing method for vulcanized rubber as described in claim 3, characterized in that, The method further includes: Real-time mold deformation monitoring data is collected during the vulcanization process to generate a process feedback log that includes displacement offset and stress anomaly threshold. Extract the abnormal fluctuation features from the process feedback log, and perform pattern matching between the abnormal fluctuation features and the historical vulcanization case library to generate process adjustment instructions; The constraint parameters of the process parameter optimization model are updated based on the process adjustment command; The updated constraint parameters are injected into the feature cross network, and the energy threshold range in the multi-scale fusion feature vector is recalculated.

5. The intelligent processing method for vulcanized rubber as described in claim 4, characterized in that, The step of extracting abnormal fluctuation features from the process feedback log and performing pattern matching between these features and a historical vulcanization case library to generate process adjustment instructions includes: The abnormal fluctuation characteristics are processed by time slicing to generate multiple fluctuation event segments and their corresponding temperature and pressure joint curves; The pre-trained anomaly attribution model is invoked to perform root cause analysis on each fluctuation event segment, generating a set of classification labels including uneven filler distribution, sudden temperature change, and mold aging. Retrieve process correction templates that match the classification tag set from the historical vulcanization case library to generate a candidate adjustment strategy set; Based on the similarity ranking between the temperature-pressure joint curve and the candidate adjustment strategy set, the strategy with the highest matching degree is selected to generate the process adjustment instruction.

6. The intelligent processing method for vulcanized rubber as described in claim 5, characterized in that, The pre-trained anomaly attribution model is invoked to perform root cause analysis on each fluctuation event segment, generating a set of classification labels including uneven filler distribution, sudden temperature changes, and mold aging, including: The fluctuation event segments are processed by keyframe extraction to generate a stress mutation frame sequence and timestamp index; Calculate the difference in thermal expansion coefficients between adjacent frames in the stress abrupt change frame sequence to generate a material deformation acceleration vector; The material deformation acceleration vector is input into the multipath convolutional subnet of the anomaly attribution model for local feature extraction, generating a spatial anomaly response map. The timestamp index is subjected to window sliding processing to generate a histogram of time-dimensional anomaly distribution; The spatial anomaly response map and the temporal dimension anomaly distribution histogram are subjected to feature cross-validation to generate a root cause probability distribution matrix. Extract the peak indices of the probability of uneven packing distribution, the probability of sudden temperature change, and the probability of mold aging from the root cause probability distribution matrix. The preset probability threshold is dynamically compared based on the peak index to generate a set of classification labels with weighted coefficients.

7. The intelligent processing method for vulcanized rubber as described in claim 2, characterized in that, The method further includes: Virtual operating condition disturbance parameters are injected before the vulcanization process is started. These virtual operating condition disturbance parameters are used to simulate pressure pulsation and temperature drift scenarios. Monitor the stability processing results of the process parameter optimization model on the disturbed operating conditions, and generate fault tolerance evaluation indicators; When the fault tolerance evaluation index is lower than the preset threshold, the parameter adaptive update mechanism of the feature cross network is triggered; The convolutional kernel weights of the feature cross network are iteratively optimized based on the difference in physical property parameters before and after the perturbation.

8. The intelligent processing method for vulcanized rubber as described in claim 7, characterized in that, The iterative optimization of the convolutional kernel weights of the feature cross network based on the difference in physical property parameters before and after the perturbation includes: Tensor alignment is performed on the difference data of physical property parameters before and after the disturbance to generate a physical property offset matrix; Extract the rate of change of thermal conductivity and the packing distribution offset vector from the property offset matrix; The rate of change of thermal conductivity coefficient and the packing distribution offset vector are concatenated to generate the physical property loss function input tensor. The prediction error gradient of the feature cross network is calculated based on the input tensor of the physical property loss function. The prediction error gradient is smoothed by momentum to generate the convolution kernel weight update direction vector; The convolution kernel parameters in the feature cross network are adjusted based on the convolution kernel weight update direction vector.

9. The intelligent processing method for vulcanized rubber as described in claim 1, characterized in that, The method further includes: Build a cross-equipment process adapter and analyze the differences in control protocols among different vulcanization equipment; The vulcanization process control instruction set is converted into a sequence of low-level drive instructions executable by the target device; The phase synchronization relationship and exception handling logic of the timing operation sequence are preserved during the conversion process; Inject real-time response parameters that match the target device to generate device-independent process control bytecode.

10. An intelligent processing system for vulcanized rubber, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 9.

Images