[Abstract] This paper mainly introduces the design and implementation of a two-stage model control system for a medium-thick plate straightening machine. The model uses a neural network as the design carrier and employs the radial basis function approximation method for calculation and analysis, which ensures the accuracy and precision of the model strategy and improves the straightening quality of the steel plate.
[Keywords] L2 model, neural network, radial basis function data table
Foreword
Establishing an accurate mathematical model has a good effect on dealing with nonlinear problems. The design of the second-level model of the straightening machine is to establish a neural network control system. Based on the approximation nonlinear function operation method of radial basis function, a reliable straightening strategy is obtained. The straightening force, torque, speed and other parameters of the straightening machine are comprehensively analyzed and set. Under the premise of ensuring that the limit value is not exceeded, the desired plastic strain ratio and flatness are obtained.
Composition and main functions of the L2 model
The Level 2 model consists of three parts: First, the Level 2 server: This computer runs Level 2. Therefore, tracking, preset value calculation, database storage, and communication are all completed by this server, which is the command center of Level 2; Second, the Human-Machine Interface (HMI): It is responsible for the application of the human-machine interface and the actual operation of the system through it. The application software is Visual Basic; Third, the Engineering Station Computer: It is used for program control and administrator maintenance. The program design is implemented in C++ and VAI environment.
The main functions include: input of PDI (basic data); output of PDO basic data; operation implementation of HMI human-machine interface; implementation of secondary database; data communication with L3; and communication with other related devices.
2. Structure and Calculation of Neuron Networks
The L2 model of the straightening machine is operated using a single-layer radial basis function network. Radial basis function (RBF) network is a novel and effective feedforward neural network. It is a linear combination of a set of radial basis functions, which has high operation speed and extrapolation ability. At the same time, it has a strong ability to approximate nonlinear functions, and the relationship between the approximation accuracy and the network structure is plastic, that is, the network can achieve the best approximation accuracy by adjusting its own structure.
Radial basis function (RBF) networks are a common type of artificial neural network, proposed by J. Moody and C.D. Arken in 1989. They use radial basis functions, such as Gaussian functions, as neuron transfer functions to approximate nonlinear functions. This network can be represented by a structure diagram (Figure 1). As shown in the figure, it consists of an input layer, a hidden layer, and an output layer. The input layer consists of input sample nodes; the hidden layers are radial basis functions, and their interaction functions are radial basis functions. For each input sample, each radial basis function in the hidden layer has an output, and the output of the hidden layer is a weighted sum of these outputs.
(Figure 1: Three-layer neuron network structure diagram)
The basic principle of radial basis function networks (RBF networks) is to fit a portion of the dependent variable to a local region using a radial basis function. Different basis functions correspond to different local regions, and the weighted sum of the entire set of basis functions forms a total function, which we consider as an approximation of a certain function. Using the Mood and Darken algorithm, this algorithm first performs cluster analysis on the input portion of the samples, usually using the k-means method, dividing the input portion into several classes. The centroid of each class is calculated as the center of the radial basis functions. Then, the width parameter of the radial basis functions is specified, and the combination coefficients of each radial basis function are determined using the least squares method.
To improve the data extraction capabilities of the data table, a derived method is used to extract data from the radial basis function neural network. The specific calculation method is as follows:
(1) Standardize each PDI input data i, i.e., perform cluster analysis. See Formula 1:
(2) Calculate the width parameter of the radial basis function, as shown in Formula 2:
(3) Calculate the weight ratio for each element in the database, as shown in Formula 3:
(4) The final output values of all K elements selected in the table are shown in Formula 4:
3. Design of the secondary model
The secondary model is designed to straighten hot steel plates. This model is based on two functionalities: an offline model and an online model. The offline model is a physical model that requires a lengthy computation process; its output is stored in a data table. After successful offline testing, the online model is used for application and practice, featuring geometric adaptation capabilities similar to a neural network.
3.1 Generation of Model Strategies
The model includes online and offline models. The results of the offline model are stored in a smart table and are adaptively derived based on the measurement results. The online model is also called a process simulator. This simulator can learn on its own to obtain a better plastic deformation ratio and straightness for straightening steel plates.
The online model is a linear combination of three calculation results (based on the confidence factor of each result): the offline model calculation results, the short-term project adaptation results, and the long-term project adaptation results.
The input data for the model are: steel grade, temperature, thickness, and plastic strain ratio. Using the steel grade and temperature, the model first records the steel properties in a data table, and then calculates the initial roll gap based on the results from the offline model. Depending on the changes in steel grade, temperature, and thickness, the model will adapt to the changes in roll gap and calculate the corresponding straightening force, torque, and plastic strain ratio.
The formulas for calculating straightening force and torque are as follows:
forceOff : The straightening force t calculated by the offline model ; forceOffC : The confidence factor of the straightening force; forceShort: The straightening force of the short-item adaptation result; forceShortC: The confidence factor of forceShort ; forceLong: The straightening force of the long-item adaptation result; forceLongC: The confidence factor of forceLong .
The online model verifies whether various limiting values (straightening force, torque) are exceeded. If the plastic ratio is reached and the limiting values are not exceeded, the model outputs the calculation results. When the plastic ratio is not reached but the limiting value is not exceeded, the model will reduce the roll gap and re-verify. When the plastic ratio is not reached but one limiting value is exceeded, the model reverts to the previous iteration value, sends the output results, and indicates that the plastic ratio cannot be reached.
3.2 Data Table Creation and Storage
Based on the physical properties and characteristics of steel plates, the model calculates and records the thickness distribution of the steel plate and calculates its trajectory according to the position of the straightening rollers when it enters the straightening machine. This calculation process considers all position settings and all steel plate characteristics. The calculation results are stored in a data table, mainly including the following key values as shown in Table 1:
Seq | Field | Description | Type | Length | Message len |
1 | Msg_ID | Information signal | Char | 5 | 5 |
2 | Rec_Len | Information length | Char | 5 | 5 |
3 | MPLATE_NO | Motherboard number | CHAR | 10 | 10 |
4 | SLAB_NO | slab number | CHAR | 10 | 10 |
5 | OCCR_DTIME | Date of Information | CHAR | 14 | 14 |
6 | FAC_CD | Factory Code | CHAR | 1 | 1 |
7 | Actual start temperature | Actual start temperature | Num | 4 | 4 |
8 | Actual end temperature | Actual ending temperature | Num | 5 | 5 |
9 | Start time | Start time | Char | 14 | 14 |
10 | End time | End time | Char | 14 | 14 |
11 | Plastification Ratio | compression amount | Num | 4 | 4 |
12 | Levelling Force | straightening force | Num | 4,2 | 5 |
13 | Levelling Speed | Straightening speed | Num | 3,2 | 4 |
14 | Threading Speed | bite speed | Num | 3,2 | 4 |
15 | Pass counter | Straightening channels | Num | 1 | 1 |
16 | Plate Status | Steel plate condition | Num | 1 | 1 |
3.3 Measurement Data Processing
The collected data is analyzed using statistical methods. Valid data is written into reports and applied to the model's adaptive functions. L1 sends measurement data values to L2 at 200ms intervals. This data is stored in a loop stack, continuously receiving feedback from L1. The tracking model then processes and analyzes the data. Unprocessed physical measurement data must be triggered by a change in the straightening machine's state to be effectively used. Similarly, if measurement conditions are no longer met (e.g., stable rolling speed, unstraightened steel plates within the straightening machine), data processing stops.
The measured values include: the length of the steel plate, the tension at the inlet and outlet of the straightener, the straightening force of the frame, the bending roll force, the straightening speed, the main motor power, the position of the inlet and outlet rolls, and the elongation. Once the processed data is valid, the model undergoes adaptive learning, and the measured data is stored in the PDO table.
3.4 HMI Human-Computer Interface
The HMI (Human-Machine Interface) has two main functions: first, to facilitate operators in viewing production data and information; and second, to enable the collection of steel plate information and material tracking data between upstream and downstream processes of the straightening machine (upstream is the ACC rapid cooling equipment, downstream is the cooling bed), maintenance of the L2 data table, receiving steel plate PDI information from L3, maintenance of the rolling program, data collection and storage, model self-learning function, and pre-calculation settings of the L2 model.
The HMI displays material information from L3 and tracking information from L2, including information on the next steel plate to be straightened, the steel plates currently being straightened, and the last three steel plates that have been straightened.
The L1 hot metal detector (HMD) detects the head and tail of the steel plate, but does not display the actual position and shape of the steel plate on the L2 HMI. Instead, it transmits the steel plate tracking information detected by L1 to L2 via communication. L2 calculates the number of straightening passes, reduction amount, torque, and other model strategies based on the data sent by L1 and parameters such as the speed of the field equipment. Between passes, the L2 model adaptively adjusts based on the real-time data from L1 to achieve a better straightening strategy and flatness. The HMI is shown in Figure 2.
(Figure 2: Tracking using the L2 model)
4. Summary
The application of neural network models improves the accuracy of secondary strategy analysis. The application of secondary models enhances the straightening quality of the straightening machine, enabling fully automated functions from receiving steel plate information from the tertiary stage, real-time material tracking, automatic calculation of straightening parameters, and adaptive learning. This maximizes the straightening quality of the steel plate, improves equipment safety, and avoids equipment damage caused by excessive torque or straightening force, demonstrating significant potential for widespread application.
Author: Yue Linping, female (1972-), Master's degree, senior electrical engineer, mainly engaged in the field management and maintenance of computer control and automation instruments.
Email: [email protected]
Mobile phone: 13864172994
Address: No. 21, Industrial North Road, Licheng District, Jinan City, Shandong Province, China (Jigang Medium and Heavy Plate 3500 Production Line) Postcode: 250101
