Variable curvature bending arc control method for roll bending machine

This application is based upon and claims priority to Chinese Patent Application No. 202410521415.7, filed on Apr. 28, 2024, the entire contents of which are incorporated herein by reference.

The present invention belongs to the field of industrial production control technology, and specifically relates to a variable curvature bending arc control method for a roll bending machine.

Roll-bending forming is a metal processing method that rolls a metal sheet or tube in an axial direction by using a roll bending machine. As a common metal forming process, it is widely used in aviation, rocket, ship, automobile, and construction fields.

The principle of roll-bending forming is that a metal sheet or tube is placed between the rollers of a roll bending machine and the metal material is pushed forward by a rolling shaft so that the metal sheet or tube is gradually bent into a desired shape by the rollers. The roller shaft consists of one or more metal rollers arranged as needed to adjust for flexible roll-bending forming, semi-flexible roll-bending forming, or rigid roll-bending forming. As the rollers pass through the metal material, they exert a certain force and bend the metal, meanwhile, the position and angle of the rollers shaft are adjusted to control the degree and shape of the metal bending. Conventional roll bending machines are manually controlled and often need to be formed and shaped repeatedly many times to satisfy the requirements, which is not only inefficient but also costly. In addition, the manufacturing process of variable curvature profile parts is more complicated. Currently the roll-bending forming of large-size and asymmetric section profile parts is still in the theoretical research stage, and effective equipment and mature experience are relatively scarce. However, the forming radius is difficult to control and the smooth transition of profiles with different curvature can not be realized for variable curvature roll-bending profile processing in the current stage.

An objective of the present invention is to provide a variable curvature bending arc control method for a roll bending machine, the method solves the technical problem that the forming radius of the variable curvature roll bending profile is difficult to control in the existing technology, which leads to the failure to realize the smooth transition of profiles with different curvature.

The technical scheme adopted by the present invention is the variable curvature bending arc control method for the roll bending machine, the method is implemented in accordance with the following steps:

step 1, collecting working data of a roll bending machine and pre-processing the data; step 2, determining a number of input layer nodes, hidden layer nodes and output layer nodes of a back propagation neural network according to the pre-processed data in step 1; step 3, encoding a weight value and a threshold value of the back propagation neural network into individuals in a population according to a real number encoding rule, and initializing the population according to a set population size and a random initialization method; step 4, performing an iterative operation on the population obtained in step 3 according to a genetic algorithm-particle swarm optimization (GA-PSO) algorithm rule to obtain a global optimal individual; step 5, decoding the global optimal individual obtained in step 4, restoring back propagation neural network structure information, and substituting into the back propagation neural network to obtain an initial weight value and a threshold value of the back propagation neural network; step 6, repeatedly learning and training the back propagation neural network according to a preset gradient descent supervised learning method and an error back propagation learning algorithm to obtain an optimized back propagation neural network; step 7, applying a back propagation neural network model optimized in step 6 to an actual control system, so as to accurately control a variable curvature bending arc of a roll-bending profile. The present invention is also characterized by, the pre-processing in step 1 is specifically as follows: collecting working data of the roll bending machine during a certain time period as a training sample, the training sample including a torque To, a roller pressure Pr, a roller position Pot, a feed velocity V and a material temperature Te, and predicting a torque value and a roller position Pot2 of the roll bending machine during a t2 time period according to a torque value and a roller position Pot1 of the roll bending machine during a t1 time period, wherein time lengths of t1 and t2 are equal; an input vector P (Tot1, Prt1, Pot1, Vt1, Tet1) and an output vector T (Pot2) are obtained by pre-processing. Step 2 is specifically as follows: step 2. 1, the number of input layer nodes matches a feature number of the training sample, and each input layer node corresponds to one feature; step 2. 2, the number of nodes at the output layer depends on whether a problem to be solved is a classification problem or a regression problem, and when it is a classification problem, the number of nodes at the output layer is the same as a number of categories; when it is a regression problem, the number of output nodes is one; an output of a k th neuron in the output layer is:

performing a decoding operation of an inverse process according to the encoding rule, recovering structure information of the back propagation neural network, including the number of nodes of an input layer, a hidden layer and an output layer and connection mode information, and recombining a decoded weight value and a decoded threshold value into a corresponding weight matrix and a threshold value vector; a process of outputting error back propagation: obtaining an output value of the back propagation neural network through a forward propagation calculation process of the sample, and learning and correcting the weight value and threshold value of the back propagation neural network by using a back propagation neural network output error; an error function is:

is a first derivative of the transfer function of the output layer;

correcting the weight value of hidden layer, the weight value from a j th input to a i th output is:

is a first derivative of the transfer function of the hidden layer;

correcting the weight values of each layer of the back propagation neural network by taking an output deviation of the back propagation neural network from the output layer to the hidden layer and then to the input layer, regarding the threshold value as the connection weight value of one neuron with an input of 1 to together participate in the adjustment of the weights.

The repeated learning in step 6 is specifically as follows: repeating a process of a forward transmission of an input mode and a backward propagation of an output error continuously, and verifying whether the output error of back propagation neural network reaches a given expected accuracy value; if the output error reaches the expected accuracy value, the learning ends; otherwise, continue to cycle.

The beneficial effects of the present invention are:

in the present invention, a variable curvature bending arc control method for a roll bending machine realizes the accurate control of the variable curvature bending arc torque for the roll bending machine by combining genetic algorithm (GA), back propagation neural network (BP) and particle swarm optimization (PSO); the control performance and robustness of the system are improved by optimizing the controller parameters, so that the roll bending machine can realize the torque control stably and quickly under the condition of large curvature change; the present invention can effectively deal with the control requirements under complex working conditions, improve the production efficiency and product quality, and has a good engineering application prospect; the present invention brings a significant optimization effect for the variable curvature bending arc torque control for the roll bending machine, and solves the technical problem that the forming radius of the variable curvature roll bending profile is difficult to control in the existing technology, which leads to the failure to realize the smooth transition of profiles with different curvature, the present invention has important practical significance and application promotion value.

FIG. 1 is a flow diagram of a variable curvature bending arc control method for a roll bending machine of the present invention;

FIG. 2 is a network topology diagram of a back propagation neural network of a variable curvature bending arc control method for a roll bending machine of the present invention.

The following is a detailed description of the present invention in combination with the drawings and the specific embodiments.

As shown in FIG. 1 and FIG. 2 , the variable curvature bending arc control method for the roll bending machine disclosed in the present invention, specifically the method is implemented in accordance with the following steps:

step 1, working data of a roll bending machine is collected and the data is pre-processed, the key feature parameters are extracted through feature selection and extraction methods, so as to establish the control model; step 2, a number of input layer nodes, hidden layer nodes and output layer nodes of a back propagation neural network is determined according to the pre-processed data in step 1; step 3, a weight value and a threshold value of a back propagation neural network are encoded into individuals in a population according to a real number encoding rule, a single piece of coding information contains all the weight values and threshold values of the back propagation neural network, which also represents an individual in the population, and the population are initialized according to a set population size and a random initialization method; step 4, an iterative operation is performed on the population obtained in step 3 according to a GA-PSO algorithm rule, and the evolutionary algebra or the output error accuracy of the back propagation neural network is preset as the end condition of the algorithm, when the algorithm satisfies the end condition, the global optimal individual is obtained, the global optimal individual is an optimal structural parameter of the back propagation neural network; step 5, the global optimal individual obtained in step 4 is decoded, back propagation neural network structure information is restored, and substituted into the back propagation neural network to obtain an initial weight value and a threshold value of the back propagation neural network; step 6, the back propagation neural network is repeatedly learned and trained according to a preset gradient descent supervised learning method and an error back propagation learning algorithm to obtain an optimized back propagation neural network; step 7, the back propagation neural network model optimized in step 6 is applied to an actual control system, so as to accurately control a variable curvature bending arc of a roll-bending profile.

In the present invention, the variable curvature bending arc control method for the roll bending machine realizes the accurate control of the variable curvature bending arc torque for the roll bending machine by combining genetic algorithm, back propagation neural network and particle swarm optimization. The control performance and robustness of the system are improved by optimizing the controller parameters, so that the roll bending machine can realize the torque control stably and quickly under the condition of large curvature change. The present invention can effectively deal with the control requirements under complex working conditions, improve production efficiency and product quality, and has a good engineering application prospect. The present invention brings a significant optimization effect for the variable curvature bending arc torque control for the roll bending machine, which has important practical significance and application promotion value.

On the basis of embodiment 1, the pre-processing in step 1 is specifically as follows:

working data of the roll bending machine during a certain time period are collected as a training sample, the training sample including a torque To, a roller pressure Pr, a roller position Pot, a feed velocity V and a material temperature Te, and a torque value and a roller position Pot2 of the roll bending machine during a t2 time period are predicted according to a torque value and a roller position Pot1 of the roll bending machine during a t1 time period, wherein time lengths of t1 and t2 are equal;

an input vector P (Tot1, Prt1, Pot1, Vt1, Tet1) and an output vector T (Pot2) are obtained by pre-processing.

Further, step 2 is specifically as follows:

step 2.1, the number of input layer nodes matches a feature number of the training sample, and each input layer node corresponds to one feature, in the present invention, the number of input layer nodes of is 5; step 2.2, the number of nodes at the output layer depends on whether a problem to be solved is a classification problem or a regression problem, and when it is a classification problem, the number of nodes at the output layer is the same as a number of categories; when it is a regression problem, the number of output nodes is one. The present invention is a regression problem, and the number of output layer nodes is 1.

Further, the GA-PSO algorithm in step 4 is specifically as follows:

step 4.1, the back propagation neural network parameter is initialized, a mutation probability P m , a crossover probability P c , a maximum particle velocity V max , an inertia constant ω, learning factors c 1 , c 2 , a number of iterations epoch and an accuracy of the solution E are preset; step 4.1.1, crossover operator design, as shown in Formulas (3) and (4): a. position crossover operation:

Further, step 5 is specifically as follows:

a decoding operation of an inverse process is performed according to the encoding rule, structure information of the back propagation neural network is recovered, including the number of nodes of an input layer, a hidden layer and an output layer and connection mode information, and a decoded weight value and a decoded threshold value are recombined into a corresponding weight matrix and a threshold value vector; a process of outputting error back propagation: an output value of the back propagation neural network is obtained through a forward propagation calculation process of the sample, and a back propagation neural network output error is used to learn and correct the weight value and threshold value of the back propagation neural network; an error function is:

is a first derivative of the transfer function of the output layer;

the weight value of the hidden layer is corrected, the weight value from a j th input to a i th output is:

is a first derivative of the transfer function of the hidden layer;

the weight values of each layer of the back propagation neural network are corrected by taking an output deviation of the back propagation neural network from the output layer to the hidden layer and then to the input layer, the threshold value is regarded as the connection weight value of one neuron with an input of 1 to together participate in the adjustment of the weights.

Further, according to the preset gradient descent supervised learning method and error back propagation learning algorithm, the back propagation neural network is repeatedly trained until satisfying the algorithm end, at the end of the algorithm, an final output of the GA-PSO-BP network is obtained, the final output refers to the parameters and weight values obtained after the training of the back propagation neural network, these parameters and weight values are the key factors to control the behavior of the model. The output result is a mapping result corresponding to the input data of the back propagation neural network. The repeated learning in step 6 is specifically as follows: a process of a forward transmission of an input mode and a backward propagation of an output error is repeated continuously, in order to make the output error of the back propagation neural network small enough, each set of training samples of the back propagation neural network usually go through hundreds of repeated training processes before the back propagation neural network can master the mapping relationship between the input and output data, every cycle memory training needs to be verified whether the output error of back propagation neural network reaches a given expected accuracy value; if the output error reaches the expected accuracy value, the learning ends; otherwise, continue to cycle.

Since the back propagation neural network model has five input parameters, that is, torque, roller pressure, roller position, feed velocity and material temperature, and one output parameter is the torque value of the roll bending machine at a next moment, the input layer has five nodes and the output layer has one node, the number of hidden layer nodes in step 2.3 is determined according to the empirical formula n=√{square root over (n i +n o )}+a;

where: n is a number of hidden layer nodes, n i is a number of input layer nodes, n o is a number of output layer nodes, a is a constant between 1-10, a value range of n is 4-13, in the embodiment, the number of hidden layer nodes is 10.