A multi-time-scale intelligent control method for steel production equipment

Abstract

The present invention relates to a multi-time-scale intelligent control method for steel production equipment, belonging to the technical field of intelligent control of steel production processes. The method comprises: collecting and preprocessing historical multimodal data to form multimodal time series data; constructing a device digital twin that integrates a physical model and a data-driven model to predict device responses, and training the device digital twin using the multimodal time series data; collecting real-time multimodal data, preprocessing it, and inputting it into the trained device digital twin to obtain predicted device responses; based on the predicted device responses, performing multi-time-scale hierarchical optimization control of control parameters using an improved multi-objective particle swarm optimization algorithm and a constrained Bayesian optimization algorithm; and adjusting the device control parameters via a PLC actuator based on the optimized control results, and feeding the adjusted device state variables back to sensors to achieve closed-loop control.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent control of steel production processes and relates to a multi-time scale intelligent control method for steel production equipment. Background Art

[0002] During steel production, the operating status and maintenance requirements of equipment change over time, making traditional equipment control methods difficult to adapt to the degradation characteristics of equipment and complex production conditions. Digital twin technology can accurately simulate the physical state of equipment and establish predictive models based on large amounts of data, but existing methods lack flexible control strategies to cope with dynamically changing production environments. Model-based intelligent optimization methods can achieve dynamic adjustment of equipment parameters, but they fail to effectively combine multi-dimensional degradation modeling with cross-production line knowledge sharing. Therefore, an intelligent control method that integrates digital twins, intelligent optimization, and adaptive control is needed to improve the accuracy and stability of equipment control. Summary of the Invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a multi-time-scale intelligent control method for steel production equipment, which can realize autonomous perception, accurate prediction and hierarchical optimization of key control parameters of the equipment, thereby improving control accuracy and operation stability.

[0004] The present invention provides a multi-time-scale intelligent control method for steel production equipment, comprising:

[0005] Step 1: Collect historical multimodal data and preprocess them to form multimodal time series data;

[0006] Step 2: Build a digital twin of the device that integrates the physical model and the data-driven model, predict the device response, and train the digital twin using multimodal time series data.

[0007] Step 3: Collect real-time multimodal data, pre-process it, and input it into the trained device digital twin to obtain the predicted value of the device response;

[0008] Step 4: Based on the predicted value of the equipment response, the control parameters are optimized at multiple time scales and hierarchically by using the improved multi-objective particle swarm optimization algorithm and constrained Bayesian optimization algorithm;

[0009] Step 5: Based on the optimized control results of step 4, the control parameters of the equipment are adjusted through the PLC actuator, and the adjusted equipment state variables are fed back to the sensor to achieve closed-loop control.

[0010] The multi-time-scale intelligent control method for steel production equipment of the present invention has the following beneficial effects:

[0011] The control method proposed in this paper constructs a digital twin of the equipment based on multimodal data, integrating a physical model and LSTM to improve prediction accuracy. By introducing a dual-time-scale optimization algorithm, it achieves a fusion of long-term and short-term dynamic control. This method addresses the slow response and adjustment difficulties of traditional steel production equipment control systems, dynamically optimizing equipment control parameters based on real-time changes during the production process. In particular, the combination of digital twin technology, optimization algorithms, and multimodal sensing enhances the system's adaptability, flexibility, and real-time responsiveness under complex production conditions. The method is highly portable and suitable for adaptive operational control of steel manufacturing equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flow chart of a multi-time-scale intelligent control method for steel production equipment of the present invention. DETAILED DESCRIPTION

[0013] The specific embodiments of the present invention are described in further detail below with reference to the accompanying drawings and embodiments. In this embodiment, the equipment object is the F1-F7 stands of a 2250mm hot continuous finishing mill.

[0014] like Figure 1 As shown, a multi-time-scale intelligent control method for steel production equipment of the present invention includes:

[0015] Step 1: Collect historical multimodal data and preprocess it to form multimodal time series data, specifically:

[0016] Step 1.1: Deploy a multimodal sensor array at key locations of the target steel production equipment to collect historical state variables, including: furnace temperature in the heating furnace process; rolling force and bending force in the roughing process; rolling force, bending force, inlet speed, outlet speed, strip inlet temperature, strip outlet temperature, stand vibration acceleration, inter-stand tension, and cooling water flow rate of all stands in the hot rolling process; and strip inlet thickness and strip outlet thickness for each process.

[0017] In this embodiment, the sensor deployment is as follows: each infrared temperature measurement array consists of 10 infrared sensors, which monitor the temperature of the head, middle, and tail of the strip respectively, with a sampling period of 50ms; a pressure sensor is installed in the hydraulic cylinder of the rolling mill, with a sampling period of 50ms; an accelerometer is arranged on the transmission side and the operating side of the rolling mill, with a sampling period of 50ms; and an X-ray thickness gauge has a sampling period of 50ms. Some data are shown in Table 1:

[0018] Table 1: Partial sampling data

[0019]

[0020] Step 1.2: All sensors are connected to a unified data acquisition system, and historical state variables are cleaned to remove outliers, denoise, and normalize to ensure the consistency and reliability of multimodal time series data.

[0021] Step 1.3: Construct multimodal time series data based on the preprocessed historical state variables:

[0022]

[0023] in, Represents the sampled value of the i-th state variable at time t.

[0024] Step 2: Build a digital twin of the device that integrates the physical model and the data-driven model, predict the device response, and train the digital twin of the device using multimodal time series data.

[0025] In specific implementation, the equipment response includes the strip outlet temperature and strip outlet thickness. The physical model reflects the actual physical properties of the equipment, and the data-driven model supplements the equipment state prediction. Step 2 is specifically as follows:

[0026] Step 2.1: Establish a physical model and use the least squares fitting method based on multimodal time series data to determine the static weight coefficient of the model:

[0027]

[0028]

[0029] in, represents the predicted value of the physical model, express The weighted sum of It corresponds to The static weight coefficient of is obtained by fitting historical multimodal data using the least squares method; Express consideration The weighted sum of the rates of change, is the dynamic weight coefficient, It represents the actual value of the equipment's response under a specific state, that is, the actual value of the strip outlet temperature and the actual value of the strip outlet thickness.

[0030] Step 2.2: Build an LSTM network as a data-driven model to transform the predicted value of the physical model The actual value of the device's response in a specific state The residual between As the input for training the LSTM data-driven model, the output is the residual compensation value .

[0031] Step 2.3: Build a device digital twin using the physical model and LSTM network. The LSTM network compensates for the errors of the physical model under different working conditions by learning the historical residual sequence, so that the predicted results of the device response output by the device digital twin are closer to the actual value of the device response. The predicted results are expressed as .

[0032] The trained digital twin of the equipment is used to predict equipment responses during real-time operation. The prediction results are used as input for the subsequent Improved Multi-Objective Particle Swarm Optimization (IMOPSO) and Constrained Bayesian Optimization (CBO) algorithms to optimize the equipment's control parameters and ensure optimal operation.

[0033] In this embodiment, in order to predict the strip outlet temperature, the following physical model is constructed:

[0034] At this time, the static weight , , The dynamic weight is obtained by fitting 3500 sets of historical data (3000 sets for training and 500 sets for testing) using the least squares method. , , .

[0035] The data-driven model uses an LSTM network to compensate for the difference in physical model predictions. The input layer is the temperature prediction residual sequence of the past 60 time steps. The hidden layer is a stacked LSTM (64 units × 2 layers), the activation function is tanh, and the dropout rate is 0.2; the output layer is a residual compensation value; the learning rate is 0.001, the training cycle is 200, and the loss function is MSE.

[0036] Step 3: Collect real-time multimodal data, pre-process it, and input it into the trained device digital twin to obtain the predicted value of the device response.

[0037] Step 4: Based on the predicted value of the equipment response, the control parameters are optimized at multiple time scales and hierarchically by using the improved multi-objective particle swarm optimization algorithm and constrained Bayesian optimization algorithm. Specifically:

[0038] Step 4.1: Get the predicted value of the equipment response in hours, and use the improved multi-objective particle swarm optimization algorithm to optimize the control parameter set. Perform hourly periodic global search optimization. The control parameters in the control parameter set constitute the particles in the improved multi-objective particle swarm optimization algorithm, specifically:

[0039] Step 4.1.1: Initialize the particle swarm, where each particle represents a set of control parameters , randomly generate particle positions represents the mth particle Position in dimension n; particle velocity express The particle speed is initialized to 10% of the control parameter range.

[0040] In specific implementation, the control parameter set P represents a subset combination selected from multimodal data collected in real time that has a positive regulatory effect on equipment operation and is adjustable; such as the temperature setting of the heating furnace, the rolling speed setting, and the cooling water flow setting.

[0041] In this embodiment, the control parameters are set as:

[0042] .

[0043] in, , The value range is constrained by security Sure.

[0044] Step 4.1.2: Construct the objective function of the improved multi-objective particle swarm optimization algorithm:

[0045]

[0046] Among them, the objective function represents the prediction result of the device digital twin With actual value The error is minimal.

[0047] Step 4.1.3: Determine whether the particle is in the global optimal position by calculating the fitness function value. If it is not the global optimal position, update the particle speed and particle position according to the following formula:

[0048]

[0049]

[0050] in, Represents the static weight coefficient of the input variable; Represents the inertia weight attenuation coefficient, which is used to balance global search and local search; and represents two random factors, , controlling the balance of the search; Represents particles The best historical position in dimension n; Represents the current global best position.

[0051] In this example, the particle swarm contains 100 particles, and the initial velocity is set to 10% of the search range; the initial inertia weight is set ; Attenuation coefficient ,satisfy ; , After 40 iterations, the optimal solution was an outlet velocity of 14.2 m / s, a reduction of 23.5%, and a cooling water flow rate of 2000 L / min.

[0052] Step 4.1.4: Re-obtain the predicted value of the device response based on the device digital twin to ensure that the optimization search direction is always consistent with the actual physical response of the device.

[0053] Step 4.1.5: Determine whether to terminate the optimization process based on the set termination conditions. If the termination conditions are met, output the global optimal position of the particle as the optimal control parameter for the next large cycle. If the termination conditions are not met, return to step 4.1.2 to continue iteration.

[0054] In this example, to achieve dynamic adaptive adjustment of control parameters over long periods and ensure high-performance system operation under multiple operating conditions, an objective function for an improved multi-objective particle swarm optimization algorithm is constructed based on the prediction results output by the device digital twin. This objective function serves as a fitness evaluation metric, guiding the improved multi-objective particle swarm optimization algorithm to perform global particle swarm optimization. This process uses the device digital twin as a predictor of device behavior throughout the particle fitness calculation and update process.

[0055] Step 4.2: Obtain the predicted value of the equipment response in a minute-based cycle, and use the constrained Bayesian optimization algorithm to optimize the control parameter set. Perform minute-level local optimization, specifically:

[0056] Step 4.2.1: Construct the control objective function The proxy model:

[0057]

[0058] in, and They represent the prediction results of the device response corresponding to the control parameters in the control parameter set P when fluctuations occur. The mean and variance of .

[0059] Step 4.2.2: Create the following expected improvement expression, which measures the improvement potential of the current control parameters relative to the historical optimal control parameter set:

[0060]

[0061] in, Represents the historical minimum control objective function value; represents the probability density function of the standard normal distribution, Represents the cumulative distribution function of the standard normal distribution.

[0062] Step 4.2.3: In the device safety area In, get The corresponding control parameters when reaching the maximum , as the optimal control parameter for the next small cycle:

[0063]

[0064] in, , p min Indicates the minimum setting value of the control parameter at which fluctuation occurs, p max Indicates the maximum set value of the control parameter at which fluctuations occur.

[0065] In specific implementation, when the control parameters fluctuate, the above steps will be repeated every minute or shorter time interval to continuously optimize the control parameters through the constrained Bayesian optimization algorithm.

[0066] In this example, the control parameters fluctuated. Specifically, when the cooling water flow rate was originally 2000 L / min and the cooling water pressure increased by 0.5 bar, this affected the cooling water flow rate's control over the finishing temperature. The predicted value of the equipment response after the fluctuation, i.e., the predicted strip outlet temperature, was obtained through the equipment digital twin. The following proxy model was established:

[0067]

[0068] In the safe zone The optimal cooling water flow rate is obtained by maximizing the EI criterion:

[0069]

[0070] in, , numerically solved to get .

[0071] After optimization using the constrained Bayesian optimization algorithm, the optimal solution is an outlet velocity of 14.2 m / s, a reduction of 23.5%, and a cooling water flow rate of 1980 L / min.

[0072] Step 5: Based on the optimized control results of step 4, the control parameters of the equipment are adjusted through the PLC actuator, and the adjusted equipment state variables are fed back to the sensor to achieve closed-loop control.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the concept of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-time scale intelligent control method for steel production equipment, characterized in that: include: Step 1: Collect historical multimodal data and preprocess them to form multimodal time series data; Step 2: Build a digital twin of the device that integrates the physical model and the data-driven model, predict the device response, and train the digital twin using multimodal time series data. Step 3: Collect real-time multimodal data, pre-process it, and input it into the trained device digital twin to obtain the predicted value of the device response; Step 4: Based on the predicted value of the equipment response, the control parameters are optimized at multiple time scales and hierarchically by using the improved multi-objective particle swarm optimization algorithm and constrained Bayesian optimization algorithm; Step 5: Based on the optimized control results of step 4, the control parameters of the equipment are adjusted through the PLC actuator, and the adjusted equipment state variables are fed back to the sensor to achieve closed-loop control; The step 4 is specifically as follows: Step 4.1: Get the predicted value of the equipment response in hours, and use the improved multi-objective particle swarm optimization algorithm to optimize the control parameter set. Perform hourly periodic global search optimization, and the control parameters in the control parameter set constitute the particles in the improved multi-objective particle swarm optimization algorithm; Step 4.2: Obtain the predicted value of the equipment response in a minute-by-minute period, and use the constrained Bayesian optimization algorithm to optimize the control parameter set. Perform minute-level local optimization; The step 4.1 is specifically as follows: Step 4.1.1: Initialize the particle swarm, each particle represents a set of control parameters P m ∈P, randomly generate particle position x mn represents the mth particle P m Position in dimension n; particle velocity v mn Indicates P m The particle speed is initialized to 10% of the control parameter range; Step 4.1.2: Construct the objective function of the improved multi-objective particle swarm optimization algorithm: Among them, the objective function represents the prediction result of the device digital twin The error with the actual value Y(t) is minimal; Step 4.1.3: Determine whether the particle is in the global optimal position by calculating the fitness function value. If it is not the global optimal position, update the particle speed and particle position according to the following formula: Where w represents the static weight coefficient of the input variable; γ represents the inertia weight attenuation coefficient, which is used to balance the global search and local search; r1 and r2 represent two random factors, r1, r2 ~ U (0, 1), which control the balance of the search; pbest mn Represents particle P m The best historical position in dimension n; gbest n represents the current global best position; Step 4.1.4: Re-obtain the predicted value of the device response based on the device digital twin to ensure that the optimization search direction is always consistent with the actual physical response of the device; Step 4.1.5: Determine whether to terminate the optimization process based on the set termination conditions. If the termination conditions are met, output the global optimal position of the particle as the optimal control parameter for the next large cycle. If the termination conditions are not met, return to step 4.1.2 to continue iteration. The step 4.2 is specifically as follows: Step 4.2.1: Construct a surrogate model to control the objective function F(P): Among them, μ(P) and σ 2 (P) represents the prediction results of the device response corresponding to the control parameters in the control parameter set P when the control parameters fluctuate The mean and variance of Step 4.2.2: Create the following expected improvement expression: Among them, F best represents the historical minimum control objective function value; F represents the probability density function of the standard normal distribution, and φ represents the cumulative distribution function of the standard normal distribution; Step 4.2.3: In the equipment safety zone B safe Obtain the control parameter p that maximizes EI(P) * , as the optimal control parameter for the next small cycle: Among them, B safe =[p min ,p max ], p min Indicates the minimum setting value of the control parameter at which fluctuation occurs, p max Indicates the maximum set value of the control parameter at which fluctuations occur.

2. The multi-time-scale intelligent control method for steel production equipment according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1.1: Deploy a multimodal sensor array at the target steel production equipment to collect historical state variables, including: furnace temperature in the heating furnace process; rolling force and bending force in the roughing process; rolling force, bending force, inlet speed, outlet speed, strip inlet temperature, strip outlet temperature, stand vibration acceleration, inter-stand tension, and cooling water flow rate of all stands in the hot rolling process; and strip inlet thickness and strip outlet thickness for each process. Step 1.2: Preprocess the historical state variables. Preprocessing includes data cleaning, denoising, and normalization to ensure data integrity and accuracy. Step 1.3: Construct multimodal time series data based on the preprocessed historical state variables: S(t)={x1(t),x2(t),...,x i (t)} Among them, x i (t) represents the sampling value of the i-th state variable at time t.

3. The multi-time-scale intelligent control method for steel production equipment according to claim 2, characterized in that: The equipment response includes the strip outlet temperature and the strip outlet thickness. Step 2 is specifically as follows: Step 2.1: Establish a physical model and use the least squares fitting method based on multimodal time series data to determine the static weight coefficient of the model: in, represents the predicted value of the physical model, ∑α i x i (t) represents x i The weighted sum of (t), α i is the corresponding x i The static weight coefficient of (t) is obtained by fitting historical multimodal data using the least squares method; ∑β i Consider x i (t) Weighted sum of the rate of change, β i is the dynamic weight coefficient, Y(t) represents the actual value of the equipment response, that is, the actual value of the strip outlet temperature and the actual value of the strip outlet thickness; Step 2.2: Build an LSTM network as a data-driven model to transform the predicted value of the physical model The residual error between the actual value Y(t) of the device response As the input of the training LSTM network, the output is the residual compensation value Step 2.3: Construct a device digital twin using the physical model and LSTM network. The LSTM network compensates for the errors of the physical model under different working conditions by learning the historical residual sequence. The predicted result of the device response output by the device digital twin is expressed as

Images