A method for controlling the operation of a sinking roller in producing ultra-thin strip steel

Abstract

The present invention belongs to the technical field of sinking roller production. The present invention provides a method for controlling the operation of sinking rollers for producing ultra-thin strip steel, comprising the following steps: by deploying a multi-physics field coupling parameter monitoring system, collecting the dynamic tension-velocity collaborative data stream of the ultra-thin strip steel and the three-dimensional vibration spectrum characteristics of the sinking roller, constructing a time-frequency domain joint working condition feature library, and using an improved support vector machine integrated learning framework to perform a probabilistic prediction of whether the current working condition is a transient working condition; if the prediction determines that it is a transient working condition, the key data features of the sinking roller shaft are extracted, and a comprehensive risk assessment model is constructed based on a random forest model. The monitoring stage focuses on the layout of key sensors, improves the accuracy of transient working condition monitoring, reduces control lag caused by data redundancy, and reduces the risk of strip breakage. Through time-frequency domain feature fusion and integrated learning, the nonlinear characteristics of transient working conditions are accurately captured, and the problem of lag in transient fluctuation prediction is optimized.

Description

Technical Field

[0001] The invention belongs to the technical field of sinking roller production, in particular to a method for controlling the operation of a sinking roller for producing ultra-thin strip steel. Background Art

[0002] Ultra-thin strip steel, also known as ultra-thin strip or ultra-thin steel strip, is a high-end category of steel materials. Its thickness is usually between 0.05mm and 0.2mm (some high-end products can reach 0.015mm, such as TISCO's "hand-torn steel"), and its width can reach more than 600mm.

[0003] The core role of the sinking roller in ultra-thin strip production is reflected in steering and friction control. It is completely driven by the friction between the strip and the roller surface. The surface grooves are designed to drain the zinc liquid, ensure the strip fits tightly, and prevent slipping or deviation.

[0004] The passive drive characteristics of the sinking roller make it extremely sensitive to tension fluctuations, especially under transient conditions. For example, during startup and shutdown, sudden speed changes, or specification switching, tension fluctuations may cause unstable rotation of the sinking roller, increasing the risk of strip vibration or strip breakage. At the same time, tension fluctuations will be amplified under transient conditions. Ultra-thin strip is more sensitive to tension changes due to its small thickness and low rigidity, making it more prone to deformation or strip breakage.

[0005] To this end, the present invention provides a method for controlling the operation of a sinking roller for producing ultra-thin strip steel. Summary of the Invention

[0006] In order to make up for the deficiencies of the prior art, at least one technical problem raised in the background technology is solved.

[0007] The technical solution adopted by the present invention to solve the technical problem is: comprising the following steps:

[0008] By deploying a multi-physics field coupling parameter monitoring system, the dynamic tension-velocity coordinated data stream of ultra-thin strip steel and the three-dimensional vibration spectrum characteristics of the sinking roller are collected. A joint working condition feature library in the time-frequency domain is constructed. An improved support vector machine integrated learning framework is used to probabilistically predict whether the current working condition is a transient condition.

[0009] If the prediction is a transient condition, the key data features of the sunken roller are extracted, and the risk level of the transient condition is classified based on the comprehensive risk assessment model constructed based on the random forest model;

[0010] The risk levels are: subcritical fluctuation, significant instability warning, and critical production interruption;

[0011] Optimization and adjustment strategies are implemented for the significant instability warning levels in different risk levels.

[0012] As a further aspect of the present invention: the deployment of the multi-physics field coupling parameter monitoring system includes a sensor layout;

[0013] For the initial layout of sensors, the importance of sensors is evaluated, and then the positions of important sensors are adjusted based on the particle swarm optimization algorithm to obtain the optimized sensor layout.

[0014] As a further solution of the present invention: the important sensor is: obtaining an important factor corresponding to the sensor;

[0015] The sensors whose important factors are greater than the important factor limit are extracted as important sensors.

[0016] As a further solution of the present invention: the process of obtaining the important factors is:

[0017] The importance of the sensor is quantified by variance contribution rate and spectrum energy concentration;

[0018] For any sensor, count the number of times the importance is greater than the importance limit, and calculate the ratio with the total number as the exceeding frequency ratio; for the case where the importance is greater than the importance limit, calculate the difference between the importance and the importance limit, take the mean of the difference, and then calculate the ratio of the mean of the difference to the importance limit as the exceeding amplitude ratio; add the exceeding frequency ratio and the exceeding amplitude ratio as the importance factor.

[0019] As a further solution of the present invention: the process of obtaining the importance is:

[0020] The ratio of the variance of each sensor data to the variance of the total data is calculated as the variance contribution rate; the vibration signal is subjected to spectral analysis, and the proportion of the dominant frequency component in the total energy is extracted as the spectrum energy concentration; the variance contribution rate and the frequency energy concentration are added together to obtain the importance.

[0021] As a further solution of the present invention: the process of using the improved support vector machine integrated learning framework to perform probability prediction on whether the current working condition is a transient working condition is as follows:

[0022] The improved SVM ensemble learning framework is trained using historical data, including data samples of transient and non-transient operating conditions; the optimal model parameters are selected through cross-validation;

[0023] The real-time sensor data is preprocessed and feature extracted to obtain time domain features, frequency domain features, and time-frequency domain joint features. These features are then fed into a trained operating condition prediction model, which outputs the probability that the operating condition in the future time window is a transient condition.

[0024] If the output probability value is greater than the probability judgment threshold, it means that the working condition in the future time window is a transient working condition.

[0025] As a further solution of the present invention: the improved SVM ensemble learning framework is a machine learning framework that combines SVM with an ensemble learning strategy;

[0026] Integrated learning strategies include Bagging integration, Boosting optimization and Stacking fusion.

[0027] As a further solution of the present invention, the risk level of the transient operating condition is divided into the following steps:

[0028] The extracted key data features are input into the comprehensive risk assessment model built based on the random forest model, and the corresponding risk value is output;

[0029] If the risk value is less than or equal to the first risk classification value, the transient operating condition is determined to be a subcritical fluctuation risk level;

[0030] If the risk value is greater than the first risk classification value and less than the second risk classification value, the transient operating condition is determined to be a significant instability warning;

[0031] If the risk value is greater than or equal to the second wind classification value, the transient operating condition is determined to be critical for production interruption.

[0032] As a further solution of the present invention, the process of executing the optimization adjustment strategy for the significant instability warning level in different risk levels is as follows:

[0033] Obtain important features and their corresponding importance scores, and optimize and adjust the important features in descending order of importance scores;

[0034] For each important feature, the ratio of its corresponding importance score to the sum of the importance scores of all important features is used as the scaling factor;

[0035] Calculate the deviation between the current value of the important feature and the target value, and multiply the deviation by the proportional factor to calculate the optimization adjustment.

[0036] As a further solution of the present invention: the process of obtaining the importance features and importance scores is as follows:

[0037] During the construction of each decision tree in the random forest, the number of times each key data feature is used for node splitting is recorded; when a key data feature is used for node splitting in the decision tree, the degree of improvement in node purity brought about by the split is calculated by the reduction in impurity;

[0038] For each key data feature, count the total amount of impurity reduced when it is used as a split node in all decision trees, and then take the average as the importance score of the feature;

[0039] The key data features corresponding to the importance score threshold are extracted as importance features.

[0040] The beneficial effects of the present invention are as follows:

[0041] 1. This invention focuses on the layout of key sensors during the monitoring phase, improves the accuracy of transient condition monitoring, reduces control lag caused by data redundancy, and lowers the risk of belt breakage. Through time-frequency domain feature fusion and integrated learning, it accurately captures the nonlinear characteristics of transient conditions and optimizes the problem of transient fluctuation prediction lag.

[0042] 2. This invention quantifies the impact of transient operating conditions on production, categorizes risk levels, and reduces misjudgments or missed judgments. It also identifies features that contribute significantly to risk assessment, improving the inability of traditional methods to assess risks in a targeted manner and providing data support for subsequent strategies.

[0043] 3. The present invention precisely implements policies for different risk levels, reduces unnecessary downtime, and improves production stability. It dynamically adjusts parameters based on feature importance, optimizes the low efficiency of the one-size-fits-all adjustment method, optimizes the tension and speed stability of ultra-thin strip steel under transient working conditions, and reduces the risk of strip breakage. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention will be further described below with reference to the accompanying drawings.

[0045] Figure 1 This is a flow chart of the steps of a method for controlling the operation of a sinking roller for producing ultra-thin strip steel according to the present invention;

[0046] Figure 2 It is a logic judgment diagram for evaluating the importance of sensors in a method for controlling the operation of a sinking roller for producing ultra-thin strip steel according to the present invention;

[0047] Figure 3 This is an architectural diagram of an operating system for controlling a sunken roller for producing ultra-thin strip steel according to the present invention. DETAILED DESCRIPTION

[0048] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0049] Example 1:

[0050] During the production of ultra-thin strip steel, the operation of the sinking roll must first predict transient operating conditions in advance, so that early warnings can be issued and optimization and adjustment strategies can be implemented to reduce the impact of transient conditions on production stability.

[0051] See also Figure 1 and Figure 3As shown, a method for controlling the operation of a sinking roller for producing ultra-thin strip steel according to an embodiment of the present invention includes the following steps:

[0052] Step S10: By deploying a multi-physics field coupling parameter monitoring system, the dynamic tension-velocity coordinated data stream of the ultra-thin strip and the three-dimensional vibration spectrum characteristics of the sinking roller are collected, and a time-frequency domain joint working condition feature library is constructed. An improved support vector machine integrated learning framework is used to perform a probabilistic prediction of whether the current working condition is a transient working condition;

[0053] In this step, the deployment of the multi-physics field coupling parameter monitoring system includes the selection of physical field sensors and sensor layout;

[0054] Physical field sensors include mechanical field detection and thermal field detection. The mechanical field includes: a dynamic tension sensor using a piezoelectric tensiometer embedded in the bearing seat of the submerged roller; a speed sensor using a non-contact laser velocimeter to monitor the strip entrance and exit speeds; a three-dimensional vibration sensor using a three-axis accelerometer installed at the end of the submerged roller shaft to measure X / Y / Z axis vibration; the thermal field includes: a platinum-rhodium thermocouple array immersed in the zinc pot at different depths (surface, middle, and bottom) to measure temperature; and an infrared thermal imager for non-contact monitoring of the strip surface temperature distribution.

[0055] For the initial layout of sensors, the importance of sensors is first evaluated, and then the positions of sensors are adjusted based on the particle swarm optimization algorithm to obtain the optimized sensor layout;

[0056] Furthermore, the importance of sensors is quantified by variance contribution rate and spectrum energy concentration, and constructed into the fitness function of the particle swarm algorithm. This drives the optimization of sensor layout towards more important areas, achieving a balance between sensor quantity and monitoring performance.

[0057] The importance of sensors is assessed based on variance contribution and spectrum energy concentration. The variance contribution is calculated as the ratio of the variance of each sensor's data to the variance of the total data. A larger variance indicates more severe fluctuations in the parameter monitored by the sensor. Spectral analysis is performed on the vibration signal, and the proportion of the dominant frequency component in the total energy is extracted as the spectrum energy concentration. A more concentrated energy indicates a location is more sensitive to specific fault modes (such as sinking roller resonance). The variance contribution and frequency energy concentration are added together to determine the importance.

[0058] The variance contribution rate quantifies the sensor's contribution to the system's global fluctuations. The larger the variance, the more dramatic the parameter changes in the area where the sensor is located. This can guide resource allocation, prioritizing the deployment of high-precision sensors in areas with large fluctuations to improve the sensitivity of anomaly detection. Spectral energy concentration uses spectrum analysis to locate frequency bands sensitive to specific modes (such as resonance). A high energy concentration indicates that the location can effectively capture the characteristic frequency of the fault.

[0059] Since the importance evaluation of sensors is calculated based on historical data of multiple productions, multiple importances are calculated for each sensor. For any sensor;

[0060] Count the number of times the importance is greater than the importance limit, and calculate the ratio with the total number as the exceeding frequency ratio; for the case where the importance is greater than the importance limit, calculate the difference between the importance and the importance limit, take the mean of the difference, and then calculate the ratio of the mean of the difference to the importance limit as the exceeding amplitude ratio; add the exceeding frequency ratio and the exceeding amplitude ratio as the importance factor; extract the sensors corresponding to the values ​​greater than the importance factor limit as important sensors.

[0061] Among them, the frequency ratio of exceeding the standard reflects the stability of the importance of the sensor in multiple data. Screening can eliminate occasionally important sensors and retain long-term key sensors. The amplitude ratio of exceeding the standard quantifies the importance of the sensor, reduces the resource occupation of marginal sensors, and ensures that the optimized layout maximizes the response capability to transient working conditions.

[0062] To adjust the positions of important sensors based on the particle swarm algorithm, the sensor position coordinates are first encoded as particle dimensions, and a particle swarm is randomly generated within the feasible domain. Each particle represents a sensor layout solution. For each particle, its comprehensive fitness function value is calculated. The particle speed is adjusted based on the individual optimal and group optimal positions, and the particle position is adjusted based on the updated speed. The process of calculating fitness, updating speed and position is repeated until the maximum number of iterations (e.g., 90) is reached or the fitness value converges. Iteration is stopped when the fitness value changes by less than a threshold after 10 consecutive iterations.

[0063] Among them, the larger the comprehensive utilization value is, the stronger the monitoring capability of the corresponding layout scheme for transient working conditions is;

[0064] For example, the axial position of the sinking roller is: x∈[0,1000]mm; the radial position of the sinking roller is: x∈[0,50]mm; the particle dimension D is: D=2N, where N is the number of important sensors. If N=6, then D=12, and each particle contains 12 coordinate values. A particle swarm containing 50 particles is randomly generated within the feasible domain, and each particle represents a sensor layout scheme. For each layout scheme, the comprehensive fitness function value is calculated. The comprehensive fitness function is the sum of the variance contribution rate and the spectral energy concentration.

[0065] The aforementioned sensor importance assessment and location optimization have at least the following benefits in controlling the operation of the sinking roll in ultra-thin strip production: Through importance assessment and screening, sensors with the greatest impact on transient operating conditions can be focused on, improving the monitoring accuracy of key parameters. Location optimization allows sensors to be placed in key areas such as the sinking roll shaft system and the strip path. This can also reduce control command lags caused by data redundancy, which can increase the risk of strip breaks.

[0066] In this step, the construction of the time-frequency domain joint working condition feature library includes: extracting time domain and frequency domain features respectively, and then fusing the time-frequency domain features;

[0067] After filtering and preprocessing the collected data, it is aligned based on the timestamp to ensure the synchronization of feature extraction;

[0068] The extraction of time domain features includes mechanical field time domain features and thermal field time domain features. The mechanical field time domain features include but are not limited to: tension fluctuation rate, velocity change rate, vibration energy; the thermal field time domain features include but are not limited to: zinc liquid temperature gradient, ultra-thin strip surface temperature uniformity;

[0069] Tension fluctuation rate: Calculate the difference between the maximum and minimum tension data, and then calculate the ratio of the difference to the mean tension data;

[0070] Rate of change: Calculate the difference between the speed value at the next moment and the speed value at the previous moment, and then calculate the ratio with the time difference between the two adjacent moments;

[0071] Vibration energy: the root mean square value of the vibration signal within the window length;

[0072] Zinc liquid temperature gradient: the temperature difference between the zinc pot surface and the bottom;

[0073] Strip surface temperature uniformity: standard deviation of the strip surface temperature;

[0074] The extraction of frequency domain features includes mechanical field frequency domain features and thermal field frequency domain features. The mechanical field frequency domain features include but are not limited to: the dominant frequency components of the vibration signal; the thermal field frequency domain features include but are not limited to: the energy of the temperature fluctuation spectrum.

[0075] Dominant frequency components of the vibration signal: Perform a fast Fourier transform (FFT) on the vibration signal to extract the spectrum amplitude and identify the dominant frequency components, such as the rotation frequency of the sinking roll and the resonance frequency of the strip;

[0076] Temperature fluctuation spectrum energy: Perform fast Fourier transform (FFT) on the zinc liquid temperature signal to extract the spectrum energy of a specific frequency band (such as 0.1-1 Hz);

[0077] Fuse and splice the time domain features with the frequency domain features to construct the time-frequency domain joint features;

[0078] Store time domain features, frequency domain features, and time-frequency domain joint features, and associate features with transient operating condition labels to form a time-frequency domain joint operating condition feature library;

[0079] In this step, an improved support vector machine ensemble learning framework is used to perform a probability prediction on whether the future working condition is a transient working condition.

[0080] The core principle of support vector machines is to map data into a high-dimensional space using kernel functions and find the optimal hyperplane to maximize the classification margin. In transient operating condition prediction, SVM can effectively handle joint time-frequency domain features and nonlinear relationships. The improved support vector machine ensemble learning framework is an advanced machine learning framework that combines SVM with ensemble learning strategies. It is used to improve the accuracy and robustness of transient operating condition prediction in ultra-thin strip production.

[0081] Among them, the construction process of the improved support vector machine ensemble learning framework is:

[0082] Kernel function optimization: Select kernel functions, including but not limited to RBF and polynomial kernels; optimize kernel parameters (such as the γ value of the RBF kernel) through grid search and cross-validation to improve model fitting ability;

[0083] For example, the vibration spectrum characteristics collected by the three-axis accelerometer at the shaft end of the submerged roller are mapped using the RBF kernel to capture the nonlinear components in the vibration signal;

[0084] Ensemble learning strategies include: Bagging integration, Boosting optimization and Stacking fusion;

[0085] Bagging ensemble: Generates multiple SVM base learners, each trained on bootstrap sampled data to reduce model variance. The base learner outputs are fused through a voting mechanism to improve prediction stability.

[0086] Boosting optimization: Using the AdaBoost algorithm, we weight misclassified samples, iteratively optimize the model, and improve the ability to identify difficult samples;

[0087] Stacking fusion: deploying multi-layer SVM models to form a hierarchical prediction structure;

[0088] For example, 50 SVM base learners were generated for bagging ensemble. Each learner was trained based on different bootstrap sampling data. The outputs of these learners were fused through majority voting, significantly improving prediction stability. During the boosting optimization process, the AdaBoost algorithm was used to weight misclassified transient operating condition samples. After 10 rounds of iterative optimization, the model's ability to identify difficult samples improved by 20%. In stacking fusion, a two-layer SVM model was deployed. The bottom model processed time domain features, and the top model integrated frequency domain features and time-frequency domain joint features.

[0089] Based on the improved support vector machine integrated learning framework and historical data, the working condition prediction model is trained. The specific process is as follows:

[0090] The improved SVM ensemble learning framework is trained using historical data. The historical data includes data samples of transient and non-transient working conditions. The data samples are: input features (time domain features, frequency domain features, and time-frequency domain joint features) and output labels (transient working condition 1 and non-transient working condition 0).

[0091] Select the optimal model parameters through cross-validation, including but not limited to: kernel function type, regularization parameter C;

[0092] The real-time sensor data is preprocessed and feature extracted to obtain time domain features, frequency domain features, and time-frequency domain joint features. These features are then fed into a trained operating condition prediction model, which outputs the probability that the operating condition in the future time window is a transient condition.

[0093] The future time window is set according to the characteristics of the process parameters (such as strip thickness and running speed), for example, it is set to 0.1s, 0.5s, etc.

[0094] The output probability value is compared with the probability judgment threshold. The probability judgment threshold is used to determine whether it is a transient operating condition. The probability judgment threshold is set by using a historical data set (including transient operating condition labels) to find the optimal threshold through grid search. For example, in the range of strip thickness 0.2-0.5mm and speed 80-120m / min, the optimal threshold is P>0.65.

[0095] If the output probability value is greater than the probability judgment threshold, it means that the working condition in the future time window is a transient working condition;

[0096] If the output probability value is less than or equal to the probability judgment threshold, it means that the working condition in the future time window is a non-transient working condition;

[0097] Step S20: If the prediction is determined to be a transient operating condition, the key data features of the sunken roller are extracted, and the risk level of the transient operating condition is classified based on the comprehensive risk assessment model constructed based on the random forest model;

[0098] Among them, the risk levels are: subcritical fluctuation, significant instability warning, and production interruption critical;

[0099] The risk of transient conditions refers to the sudden or abnormal operation of the sinking roll during the production of ultra-thin strip steel. This condition may cause fluctuations in parameters such as the tension and speed of the ultra-thin strip steel, affecting the stability of the production process and even leading to production accidents such as strip breakage. The above-mentioned risk level classification also divides the degree of impact of transient conditions on production stability.

[0100] In this step, if it is a transient working condition, the key data features of the sinking roller are extracted, and the process of constructing a comprehensive risk assessment model through the random forest model algorithm is as follows:

[0101] Extract key data features of the sinking roll under transient working conditions from the multi-physics field coupling parameter monitoring system, including but not limited to: the instantaneous speed and speed fluctuation rate of the sinking roll, the sudden change rate of strip tension, the peak value of tension fluctuation, and the temperature gradient change rate of the zinc liquid;

[0102] The extracted key data features are input into the comprehensive risk assessment model built based on the random forest model, and the corresponding risk value is output;

[0103] Among them, the comprehensive risk assessment model is a model pre-trained and constructed by researchers in this field based on the operating characteristics and technical requirements of the sinking roll in the ultra-thin strip production process using the random forest model algorithm. The model uses the key data features of the sinking roll as input and the risk value as output;

[0104] Comparing the risk value with the first risk classification value and the second risk classification value respectively;

[0105] If the risk value is less than or equal to the first risk classification value, the transient operating condition is determined to be a subcritical fluctuation risk level;

[0106] If the risk value is greater than the first risk classification value and less than the second risk classification value, the transient operating condition is determined to be a significant instability warning;

[0107] If the risk value is greater than or equal to the second wind classification value, the transient operating condition is determined to be critical for production interruption;

[0108] Optionally, the process of using the random forest model to assess risk levels also includes an assessment of the importance of key data features. The importance of a feature is quantified by the number of times it is used for splitting in the decision tree and the degree of improvement in node purity after the split.

[0109] Specifically, during the construction of each decision tree in the random forest, the number of times each key data feature is used for node splitting is recorded; the more times it is used for splitting, the more important the key data feature is in the decision-making process; when a key data feature is used for node splitting in a decision tree, the degree of improvement in node purity brought about by the split is calculated; the degree of improvement in node purity can be measured by the reduction in impurity, such as the reduction in Gini impurity or information entropy; the greater the average impurity reduction brought by a feature across all decision trees, the greater the contribution of the feature to the model's prediction results;

[0110] For each key data feature, count the total amount of impurity reduced when it is used as a split node in all decision trees, and then take the average as the importance score of the feature;

[0111] According to the pre-set importance score threshold, the key data features corresponding to the importance score threshold that is greater than the importance score threshold are extracted as importance features;

[0112] The above-mentioned evaluation of the importance scores of key data features can identify key data features that have a great impact on the prediction results, thereby enabling more targeted selection and optimization of key data features and more accurate evaluation;

[0113] Step S30: Execute different optimization strategies for different risk levels;

[0114] In this step, for transient operating conditions at the subcritical fluctuation risk level, a strategy of continuing monitoring is implemented;

[0115] For transient conditions that are critical for production interruption, a strategy to interrupt the production process is implemented;

[0116] For transient operating conditions that are in a significant instability warning state, optimize and adjust the strategy to reduce the risk of transient operating conditions to production stability.

[0117] Specifically, the process of executing the optimization and adjustment strategy is as follows:

[0118] Based on the aforementioned identified importance features and obtaining corresponding importance scores, the importance features are optimized and adjusted in descending order of importance scores;

[0119] For each important feature, the ratio of its corresponding importance score to the sum of the importance scores of all important features is used as the scaling factor;

[0120] Quantify risk contribution by feature importance, and then use proportional factors to adjust the optimal allocation of resources;

[0121] Calculate the deviation between the current value of the important feature and the target value (i.e., the preset threshold value of the running process), and multiply the deviation by the proportional factor to calculate the optimization adjustment;

[0122] After determining the optimal adjustment amount, corresponding control measures can be taken to adjust the characteristic value. For example, for the instantaneous speed of the sinking roll, the speed can be changed by adjusting the output power of the drive motor; for the strip tension, the tension can be changed by adjusting the set value of the tension control system or adjusting the position of the sinking roll.

[0123] The main technical solution of this embodiment is as follows: by deploying a multi-physics field coupling parameter monitoring system, strip tension-speed data streams and sinker roll vibration spectra are collected. After sensor importance assessment (based on variance contribution rate and spectral energy concentration) and particle swarm optimization layout, a time-frequency domain joint operating condition feature library is constructed. Then, an improved support vector machine ensemble learning framework is used to predict whether the future operating condition will be transient. If a transient condition is predicted, key data features such as the instantaneous speed of the sinker roll and the strip tension mutation rate are extracted and input into a comprehensive risk assessment model based on a random forest model. The risk value is calculated and compared with the threshold, and the risk is divided into three levels: "subcritical fluctuation", "significant instability warning", and "production interruption critical". During this process, key influencing factors are screened through feature importance assessment (such as the number of splits and node purity improvement). Corresponding strategies are implemented according to the risk level: subcritical fluctuations continue to be monitored; production is immediately interrupted when the production interruption critical is reached; and when a significant instability warning is issued, the optimization adjustment amount is calculated by multiplying the deviation value proportional factor from high to low feature importance scores to reduce the impact of the transient condition on production.

[0124] This embodiment has at least the following beneficial effects: focusing on the layout of key sensors in the monitoring phase, improving the accuracy of transient working condition monitoring, reducing control lag caused by data redundancy, and reducing the risk of belt breakage; accurately capturing the nonlinear characteristics of transient working conditions through time-frequency domain feature fusion and integrated learning, and optimizing the problem of transient fluctuation prediction lag; quantifying the impact of transient working conditions on production, realizing the division of risk levels, and reducing misjudgments or missed judgments; identifying features that contribute greatly to risk assessment, optimizing the problem that traditional methods cannot assess risks in a targeted manner, and providing data support for subsequent strategies; implementing precise policies for different risk levels, reducing unnecessary shutdowns, and improving production stability; dynamically adjusting parameters based on feature importance, optimizing the low efficiency of the one-size-fits-all adjustment method, optimizing the tension and speed stability of ultra-thin strip steel under transient working conditions, and reducing the risk of belt breakage.

[0125] Example 2:

[0126] Based on the same inventive concept as the method for controlling the operation of the sinking roller for producing ultra-thin strip steel in the above embodiment, Figure 2As shown, the present application provides an operation system for controlling a sinking roller for producing ultra-thin strip steel, wherein the system specifically includes:

[0127] Transient operating condition prediction module: By deploying a multi-physics field coupling parameter monitoring system, the dynamic tension-velocity coordinated data stream of ultra-thin strip steel and the three-dimensional vibration spectrum characteristics of the sinking roller are collected. A joint operating condition feature library in the time-frequency domain is constructed. An improved support vector machine integrated learning framework is used to probabilistically predict whether the current operating condition is a transient condition.

[0128] By deploying a multi-physics field coupled parameter monitoring system, the system collects strip tension-velocity data streams and the vibration spectrum of the submerged roller. After sensor importance assessment (based on variance contribution and spectrum energy concentration) and particle swarm optimization layout, a joint operating condition feature library in the time-frequency domain is constructed. An improved support vector machine ensemble learning framework is then used to predict whether future operating conditions will be transient.

[0129] Risk analysis module: If the prediction is a transient operating condition, the key data features of the sunken roller are extracted and the risk level of the transient operating condition is classified based on the comprehensive risk assessment model constructed based on the random forest model;

[0130] Among them, the risk levels are: subcritical fluctuation, significant instability warning, and production interruption critical;

[0131] If a transient operating condition is predicted, key data features such as the instantaneous speed of the sinking roll and the sudden change rate of strip tension are extracted and input into a comprehensive risk assessment model based on a random forest model. The risk value is calculated and compared with the threshold value, and the risk is divided into three levels: "subcritical fluctuation", "significant instability warning", and "production interruption critical". During this process, key influencing factors are screened through feature importance assessment.

[0132] Hierarchical optimization module: Execute different optimization strategies for different risk levels;

[0133] For transient conditions at the subcritical fluctuation risk level, a strategy of continued monitoring is implemented; for transient conditions at the critical level of production interruption, a strategy of interrupting the production process is implemented; for transient conditions at the significant instability warning level, an optimization and adjustment strategy is implemented to optimize the risks posed by transient conditions to production stability.

[0134] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for controlling the operation of a sinking roller for producing ultra-thin strip steel, characterized in that: The following steps are involved: By deploying a multi-physics field coupling parameter monitoring system, the dynamic tension-velocity coordinated data stream of ultra-thin strip steel and the three-dimensional vibration spectrum characteristics of the sinking roller are collected. A joint working condition feature library in the time-frequency domain is constructed. An improved support vector machine integrated learning framework is used to probabilistically predict whether the current working condition is a transient condition. The deployment of the multi-physics field coupling parameter monitoring system includes sensor layout; For the initial layout of sensors, the importance of sensors is evaluated, and then the positions of important sensors are adjusted based on the particle swarm optimization algorithm to obtain the optimized sensor layout; The importance of the sensor is quantified by variance contribution rate and spectrum energy concentration; For any sensor, count the number of times the importance is greater than the importance limit and calculate the ratio with the total number as the frequency ratio of exceeding the standard; if the importance is greater than the importance limit, calculate the difference between the importance and the importance limit, take the mean of the difference, and then calculate the ratio of the mean of the difference to the importance limit as the amplitude ratio of exceeding the standard; The excess frequency ratio and the excess amplitude ratio are added together as the important factor; Extract sensors whose important factors are greater than the important factor limit as important sensors; The improved support vector machine ensemble learning framework is a machine learning framework that combines SVM with an ensemble learning strategy; Ensemble learning strategies include bagging integration, boosting optimization and stacking fusion; If the prediction is a transient condition, the key data features of the sunken roller are extracted, and the risk level of the transient condition is classified based on the comprehensive risk assessment model constructed based on the random forest model; The risk levels are: subcritical fluctuation, significant instability warning, and critical production interruption; The extracted key data features are input into the comprehensive risk assessment model built based on the random forest model, and the corresponding risk value is output; If the risk value is less than or equal to the first risk classification value, the transient operating condition is determined to be a subcritical fluctuation risk level; If the risk value is greater than the first risk classification value and less than the second risk classification value, the transient operating condition is determined to be a significant instability warning; If the risk value is greater than or equal to the second wind classification value, the transient operating condition is determined to be critical for production interruption; Optimization and adjustment strategies are implemented for the significant instability warning levels in different risk levels.

2. The method for controlling the operation of a sinking roller for producing ultra-thin strip steel according to claim 1, characterized in that: The process of obtaining the importance is as follows: The ratio of the variance of each sensor data to the variance of the total data is calculated as the variance contribution rate; the vibration signal is subjected to spectral analysis, and the proportion of the dominant frequency component in the total energy is extracted as the spectrum energy concentration; the variance contribution rate and the frequency energy concentration are added together to obtain the importance.

3. The method for controlling the operation of a sinking roller for producing ultra-thin strip steel according to claim 1, characterized in that: The process of probabilistically predicting whether the current operating condition is a transient operating condition using the improved support vector machine ensemble learning framework is as follows: The improved SVM ensemble learning framework is trained using historical data, including data samples of transient and non-transient operating conditions; the optimal model parameters are selected through cross-validation; The real-time sensor data is preprocessed and feature extracted to obtain time domain features, frequency domain features, and time-frequency domain joint features. These features are then fed into a trained operating condition prediction model, which outputs the probability that the operating condition in the future time window is a transient condition. If the output probability value is greater than the probability judgment threshold, it means that the working condition in the future time window is a transient working condition.

4. The method for controlling the operation of a sinking roller for producing ultra-thin strip steel according to claim 1, characterized in that: The process of executing the optimization adjustment strategy for the significant instability warning level in different risk levels is as follows: Obtain important features and their corresponding importance scores, and optimize and adjust the important features in descending order of importance scores; For each important feature, the ratio of its corresponding importance score to the sum of the importance scores of all important features is used as the scaling factor; Calculate the deviation between the current value of the important feature and the target value, and multiply the deviation by the proportional factor to calculate the optimization adjustment.

5. The method for controlling the operation of a sinking roller for producing ultra-thin strip steel according to claim 4, characterized in that: The process of obtaining the importance features and importance scores is as follows: During the construction of each decision tree in the random forest, the number of times each key data feature is used for node splitting is recorded; when a key data feature is used for node splitting in the decision tree, the degree of improvement in node purity brought about by the split is calculated by the reduction in impurity; For each key data feature, count the total amount of impurity reduced when it is used as a split node in all decision trees, and then take the average as the importance score of the feature; The key data features corresponding to the importance score threshold are extracted as importance features.

Images