I am using a PID Neural Network (PIDNN) algorithm to control a multi-zoned temperature hardware. The aim of the algorithm is to minimize the error with the Backpropagation rule and weights updating.

The device is commanded by B&R X20CP1484 CPU.

Unfortunately, the device has a very large inertia (SetTemperature 60°, ActualTemperature can reach 72°). Stabilizing the system requires a very long time, but always with few degrees above and below. PS: This is the case when controlling only one zone (the model has 4 zones, let alone controlling the 4 zones).

I thought about decoupling, and added all necessary equations. Still, the system has a very large inertia, heating is unstoppable. I am using a PWM too, so when the sensor value is equal to set value, duty is 0. Then temperature keeps rising even when the heater is off.

The question is: Are the PID parameters wrong? Is there a problem with the PWM period?

The same model was controlled before with GPC and had perfect results.

Device is using 4 SSR, the 4 sensors are few CMs close to the thermocouple.

Any suggestions? Thanks in advance.

  • \$\begingroup\$ Hello Med, take a look at Google's how to tune PID controller for heating for some typical ways to set PID. \$\endgroup\$ – rdtsc Oct 9 '18 at 10:16
  • \$\begingroup\$ @rdtsc Thank you for your reply , the theoretical modeling result is satisfying, I am trying to see the result in real time hardware model , so the point is not actually tuning the PID to get perfect parameters , the aim of the Backpropagation is to minimize the error , minimizing the error will automatically lead to good parameters \$\endgroup\$ – Med Chaouechi Oct 9 '18 at 10:22
  • \$\begingroup\$ @Med - I've split the text where you had used 1 x Enter (here we need 2 x Enter for a new paragraph), as the text was very difficult to read without any "white space". However, although your English is good, in a few places I had to guess where some of your sentences started & ended. Therefore please can you check that I didn't alter the meaning anywhere? You can make your own edits to this (hopefully improved) version, if needed. Thanks. \$\endgroup\$ – SamGibson Oct 9 '18 at 11:35
  • \$\begingroup\$ Why do you expect the neural network to magically tolerate arbitrarily large delays? What, in the neural network, implements the "integration" behavior? The basic network should trivially implement the "proportional" behavior. Even if the input data is adequately clean, then the "differential" behavior still requires an awareness of past-history; how are you implementing that? \$\endgroup\$ – analogsystemsrf Feb 11 '19 at 14:57
  • \$\begingroup\$ @analogsystemsrf , thank you for your reply, I was wondering about the accuracy of the PID parameters obtained from the neural network output, either the problem is related to the learning process or maybe somewhere else (bad choice of activation function, of leartning rate, random initial weights). I wasn't expecting anything to magically happen, as it is my first project using the neural networks, I wanted to make sure the process on a software scale is done correctly, the problem might be hardware-related, GPC was implemented on the same hardware without any remarkable delays. Thank you! \$\endgroup\$ – Med Chaouechi Feb 13 '19 at 5:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.