Straight line laser cutting intelligent correction method

Abstract

The present application relates to a kind of linear laser cutting intelligent deviation correction method, comprising the following steps, S1, deviation correction network is constructed;S2, deviation correction network is trained;S3, deviation correction is carried out using deviation correction network;S4, deviation monitoring, whether deviation is controllable is judged.The present application uses neural network to fit the interference between deviation corrector and cutting machine, so that deviation corrector has predictive function to this interference, and through the method of neural network with predictive function and expert control to control deviation corrector, to eliminate the influence caused by the interference, to replace the function of another deviation corrector, so as to save a deviation corrector, while the same effect as using 2 deviation correctors can be achieved;The number of deviation correctors required in the deviation correction scheme of traditional linear laser cutting device is reduced from 2 to 1, not only saving cost, but also ensuring cutting quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neural network application, and particularly relates to a linear laser cutting intelligent deviation correction method. BACKGROUND

[0002] At present, the deviation correction scheme of the linear laser cutting device on the market adopts almost all the traditional PID control strategy, and the device principle composition is as shown in the figure: Figure 1 The to-be-cut strip passes through the transmission roller and moves, and when moving to the laser cutting device with a fixed posture, the strip is penetrated by laser at the cutting point and is cut into two halves along the transmission direction with the continuous transmission movement of the strip; when the strip passes through the cutting point, the strip is cut into left and right parts, and the widths of the two parts are obtained by the monitoring device installed near the laser cutting device, which are WcL and WcR respectively; in order to make WcL=WcR, two deviation correctors are used, the deviation correctors move left and right to push the strip in the horizontal direction according to the deviation (i.e. |WcL-WcR|) obtained by the monitoring device, so that the strip is offset in the horizontal direction, thereby achieving the purpose of deviation correction; the deviation corrector 1 is installed at a distance d from the laser cutting device, as the outer ring in the PID control, mainly used to eliminate the steady-state deviation; the deviation corrector 2 is installed near the laser cutting device, as the inner ring in the PID control, used to eliminate the oscillation deviation and random disturbance; some deviation correction schemes are simplified on the basis of the traditional PID control strategy, and only the inner ring or the outer ring is used to control the deviation correction, that is, only the deviation corrector 2 or only the deviation corrector 1 is used for deviation correction.

[0003] However, the traditional PID control strategy divides the deviation correction control system into inner ring control and outer ring control, and uses one deviation corrector respectively, and a total of two deviation correctors are used, which increases the cost compared with the strategy of using only one deviation corrector for deviation correction. Figure 1 The strategy of using only the inner ring control, that is, using only the deviation corrector 2 in the figure, cannot effectively control the steady-state deviation because only the inner ring is used without the outer ring, so that the stability of the deviation correction is poor, thereby causing the sawtooth problem of the cutting edge. Figure 1 The strategy of using only the inner ring control, that is, using only the deviation corrector 1 in the figure, cannot control the interference existing between the deviation correctors and the cutting machine because only the outer ring is used without the inner ring, and the distance between the monitoring device and the deviation corrector causes the lag of the deviation correction action, so that the deviation correction action may not achieve the deviation correction effect at all. SUMMARY

[0004] The present application solves the technical problem that the traditional linear laser cutting device requires at least two deviation correctors to achieve good effect, and the cost is high due to the large number of deviation correctors.

[0005] The technical scheme adopted by the present application to solve its technical problems is: a linear laser cutting intelligent deviation correction method, comprising the following steps,

[0006] S1, constructing a deviation correction network;

[0007] S2, training the deviation correction network;

[0008] S3, using the deviation correction network for deviation correction;

[0009] S4, deviation monitoring, judging whether the deviation is controllable.

[0010] Further, in the step S1 of the present application, a noise fitting network is constructed, a fully connected network is used, the number of layers and the number of nodes of each layer can be adjusted, and the activation function is sigmoid function; the loss function is loss=(error-e) 2 ; error represents the deviation value detected by the monitoring device, e represents the output value of the output node of the last layer of the neural network; the deviation obtained by the monitoring device for multiple times is used as the input of the noise fitting network, the next sequence update delay is determined according to the loss value according to the expert control logic, the deviation sequence is updated according to the next sequence update delay, which is used as the next input of the noise fitting network, and the above method is iterated; the noise fitting network is trained by updating the input sequence until multiple times loss 2 , that is, the deviation obtained by the monitoring device.

[0011] Further, in the step S2 of the present application, the part of the deviation correction network corresponding to the noise fitting network is trained in the same way as training the noise fitting network, and in the training process, the output c of the neural network and the basic deviation correction amount offset obtained by the expert control logic are brought into the formula corr as the current deviation correction amount of the deviation corrector, which is transmitted to the deviation corrector and controls the deviation corrector to perform the deviation correction action.

[0012] Further, in the step S3 of the present application, the trained deviation correction network stops reverse updating of the weight, only forward propagation is performed, the deviation correction amount is calculated using c and offset obtained by each iteration, which is transmitted to the deviation corrector and controls the deviation corrector to perform the deviation correction action.

[0013] Further, in step S4, the loss of the correction network is monitored during correction using the correction network, i.e. the deviation value obtained by the monitoring device, and if the deviation value exceeds a set threshold value for a set number of times in succession, a noise fitting network is re-built and trained according to the noise fitting network structure in step S1, and after successful training, the new noise fitting network is used to build the correction system according to the steps after successful training of the noise fitting network.

[0014] The present application has the advantages of solving the defects in the background art, using a neural network to fit Figure 1 the interference between the middle correction device 1 and the cutting machine, so that the correction device 1 has a predictive function for the interference, and the correction device is controlled by a neural network with a predictive function combined with an expert control method to eliminate the influence caused by the interference, thereby replacing the function of the correction device 2, thereby saving one correction device while achieving the same effect as using two correction devices; reducing the number of correction devices required for the correction scheme in the traditional straight-line laser cutting device from two to one, not only saving costs, but also ensuring cutting quality. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a schematic diagram of a traditional straight-line laser cutting device;

[0016] Figure 2 is a flowchart of the correction method of the present application;

[0017] Figure 3 is a flowchart of the correction method of the present application with added feasibility verification;

[0018] Figure 4 is a structural diagram of the noise fitting network of the present application;

[0019] Figure 5 is a network structure diagram of the correction network of the present application;

[0020] Figure 6 is a schematic diagram of the expert control logic of the present application. DETAILED DESCRIPTION

[0021] The present application will now be further described in detail in conjunction with the accompanying drawings and preferred embodiments. These drawings are simplified schematic diagrams and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0022] As Figures 1-6The illustrated linear laser cutting intelligent correction method uses a neural network to verify the feasibility of the method, and after verification, the neural network is used again to fit the rule between the cutting deviation and the correction action of the corrector, and then the neural network containing the rule is combined with the expert control logic to control one corrector, thereby achieving the effect of two correctors.

[0023] As Figure 2 shown, the specific steps are as follows:

[0024] 1. Construct a correction network

[0025] A noise fitting network is constructed, and the structure of the noise fitting network is as shown in Figure 4 A fully connected network is used, and the number of layers and the number of nodes of each layer can be adjusted based on experiments. In the present application, it is a 7-layer fully connected layer, and the number of nodes of each layer is 32, 64, 128, 64, 32, 16, and 1, respectively. The activation function is a sigmoid function.

[0026] Wherein, f(1), f(2)…f(32) represent the deviation values detected by the monitoring device for the first, second, …, 32th times;

[0027] The calculation method of the deviation value is WcL-WcR, and the units of WcL and WcR are both mm.

[0028] The loss function is loss=(error-e) 2 .

[0029] Error represents the deviation value detected by the monitoring device at the current time, i.e. error=f(33), and e represents the output value of the output node of the last layer of the neural network.

[0030] Sequence update description:

[0031] Sequence update means that the deviation f(33) transmitted by the next monitoring device is used to update the original sequence f(1), f(2)…f(32) in a pipeline, i.e.

[0032] f(i)=f(i+1),i=1,2…32;

[0033] If the update delay is n, it means that after n times of deviation transmitted by the monitoring device, the sequence is updated in a pipeline, i.e.

[0034] f(i)=f(i+n),i=1,2…32;

[0035] The deviations obtained by the monitoring device for 32 consecutive times are used as the inputs of the noise fitting network, i.e. Figure 4 f(1), f2(2)…f(32) in

[0036] According to the loss value, the next sequence update delay n is determined according to the expert control logic, the bias sequence is updated according to n, as the next input of the noise fitting network, and the above method is iterated;

[0037] Wherein the expert control logic is as shown in Figure 6 , wherein n represents the next sequence update delay, and offset is the basic offset of the corrector in the subsequent step;

[0038] By updating the input sequence, the noise fitting network is trained until 64 consecutive times loss < tv, wherein tv is a verification loss threshold, and in the present application, the value is 0.002; the noise fitting network training is successful.

[0039] After the noise fitting network is trained successfully, as shown in Figure 5 , the noise fitting network is added Figure 5 The part that is more than the output layer is added to the neuron, and the output c of the neuron is used as an additional input in the input layer to form a closed loop.

[0040] The weight parameters of the trained noise fitting network are used as the weight parameters of the corresponding part of the correction network.

[0041] The loss function of the correction network is loss = error 2 , that is, the deviation obtained by the current monitoring device;

[0042] 2, training the correction network

[0043] The part corresponding to the noise fitting network of the correction network is trained according to the same method as training the noise fitting network. In the training process, the output c of the neural network and the basic offset offset obtained by the expert control logic are brought into the formula to calculate corr as the current correction amount of the corrector, which is transmitted to the corrector and controls the corrector to perform the correction action.

[0044] The correction network is trained according to the method of the step until 1000 consecutive loss loss < tc, or the training iteration reaches max_epoch.

[0045] Wherein, tc represents the loss threshold, which is 0.004 in the present application, and max_epoch represents the maximum iteration training number, which is 1000 in the present application.

[0046] 3, using the correction network to correct

[0047] The trained rectification network stops reverse updating of the weight and only performs forward propagation, and the iteration method is the same as that in step 2, the offset obtained in each iteration is used to calculate the rectification amount, which is transmitted to the rectifier and used to control the rectifier to perform rectification, and the rectification amount calculation formula is the same as that in step 2;

[0048] So far, an intelligent rectification system composed of a rectifier, a monitoring device, a rectification network and expert control logic has been built, which can be used for linear laser cutting rectification tasks.

[0049] 4, deviation monitoring

[0050] In the process of rectification using the rectification network, the loss of the rectification network is monitored, that is, the deviation value obtained by the monitoring device, if 25 of the 100 consecutive deviations exceed tc (tc has the same meaning as tc in step 4), a noise fitting network is rebuilt and trained according to the noise fitting network structure in step 1, and after the training is successful, the new noise fitting network is used to build the rectification system according to the steps after the noise fitting network is trained successfully.

[0051] It should be noted that:

[0052] The above steps are based on the premise that the noise fitting network is trained successfully, if the iteration upper limit is exceeded and the loss <t is still not achieved for 64 consecutive times, it means that the method of the present application is not feasible under the current environment, and the iteration upper limit threshold in the present application is 1500;

[0053] If the verification is not feasible, check whether the rectifier, monitoring device, overall mechanical device and the like are worn, aged or have external interference problems, if there are problems, solve the problems, and then continue to verify the feasibility, if the verification is feasible, use the new noise fitting network to build the rectification system according to the steps after the noise fitting network is trained successfully, otherwise it means that there is a difficult-to-find and solve interference, and the method of the present application is not feasible under the interference.

[0054] The above description in the specification is only a specific embodiment of the present application, and various examples do not limit the essential content of the present application, and those skilled in the art can modify or deform the previously described specific embodiments without departing from the essence and scope of the present application after reading the specification.

Claims

1. A method for intelligent deviation correction in linear laser cutting, characterized in that: Includes the following steps, S1. Construct a correction network; In step S1, a noise fitting network is constructed using a fully connected network. The number of layers and the number of nodes in each layer are adjustable, and the activation function is the sigmoid function. The loss function is loss = (error - e). 2 ;error represents the deviation value detected by the monitoring device, and 'e' represents the output value of the last layer output node of the neural network; The deviations obtained by the monitoring device continuously for multiple times are used as the input of the noise fitting network. The next sequence update delay is determined according to the expert control logic based on the loss value. The deviation sequence is updated according to the next sequence update delay and used as the next input of the noise fitting network. Iteration is performed according to the above method. By updating the input sequence, the noise fitting network is trained until the loss < tv continuously for multiple times, where tv is the verification loss threshold. After the noise fitting network is successfully trained, an additional neuron is added to the output layer of the noise fitting network, and the output c of this neuron is used as an additional input added to the input layer to form a closed loop. The weight parameters of the trained noise fitting network are used as the weight parameters of the corresponding part of the deviation correction network. The loss function of the deviation correction network is loss = error 2 , that is, the deviation obtained by the current monitoring device; S2, Training the correction network; In step S2, the part of the bias correction network corresponding to the noise fitting network is trained using the same method as the noise fitting network. During training, in each iteration, the output c of the neural network and the basic bias correction amount offset obtained from the expert control logic are substituted into the formula. The calculated corr is used as the current correction value of the correction device, which is then transmitted to the correction device and controls it to perform correction actions. S3. Use a correction network to correct the bias. Stop the backpropagation of the trained correction network and only perform forward propagation. S4. Deviation monitoring: Determine whether the deviation is controllable.

2. The intelligent deviation correction method for linear laser cutting as described in claim 1, characterized in that: In step S3, the c and offset obtained in each iteration are used to calculate the correction amount, which is then transmitted to the correction device and the correction device is controlled to perform the correction action.

3. The intelligent deviation correction method for linear laser cutting as described in claim 1, characterized in that: In step S4, during the correction process using the correction network, the loss of the correction network is monitored, i.e., the deviation value obtained by the monitoring device. If the deviation value exceeds a set threshold a certain number of times in a series of consecutive deviations, a new noise fitting network is rebuilt and trained according to the noise fitting network structure in step S1. After successful training, the correction system is built using the new noise fitting network according to the steps after the noise fitting network is successfully trained.

Images