# The term Feed-forward and its meaning?

I am trying to understand the technical meaning of the term 'feed-forward', when and where it can be used and where it cannot be used? I have seen this term in various different areas.

For example:

Vector Field oriented of AC Motors

Convolutional Neural Networks

Digital Filter Design

It looks like that this is a fundamental term. But I could not find its fundamental technical definition.

What phenomenon is encapsulated in this term. And if a certain phenomenon is labeled as 'feed-forward' then what are the attributes and properties that are understood that will be possessed by that phenomenon?

Feed forward refers to the direction of the signal flow. For feed forward, the direction is, well, forward :-)

I think it is easier to show an example. I know that many "sigma-delta" ADCs (analog to digital converters) use a combination of feedback and feed forward. The general direction of the signal flow is from left to right. The input of the ADC is at the left, the output at the right.

Note how there is a feedback line at the bottom of the diagram, from the output of the Quantizer back via b to a summation point near the input.

Note that by changing the value of b this can be either positive feedback or negative feedback. But it is feedback as the signal "goes back" (in the direction of the input).

Then there are the three signals at the top of the diagram which feed a signal from left to right so in the "forward" direction. These are the "feed forward" paths. Again, depending on the coefficients a1 and a2 these signals can give either positive feed forward or negative feed forward. But both are feed forward as the direction is towards the output.

The picture below represents a velocity controller that you can usually find in AC servo motor. It's just a fraction of the entire cascade loop.

The low frequency signals (slow response) go through lower path, this is: velocity setpoint ($$\ \Omega{set}\$$) , actual velocity ($$\ \Omega{m}\$$) from encoder feedback, lowpass filter ($$\ T_{f}\$$), PI controller ($$\ K_p, T_i\$$). The output is the torque setpoint for another PI controller. This path is used to correct the velocity error "a posteriori" , meaning the error is present and we tend to reduce it.

The higher frequencies (fast response) signals go through upper path, so called feedforward. With a known system friction ($$\F\$$) and known system inertia ($$\J\$$), with known static load torque ($$\M_{teze}\$$), with known setpoint speed and its rate of change ($$\\dfrac{d}{dt}\$$), we can predict the torque in advance "a priori". This is very fast acting path that injects the signal forward before the velocity error yet exist. Let me step back first...A control system needs an input and output, and some method to measure feedback. Open loop examples include magnetic stepper motors, while closed loop tend to be linearized for one or more variables.

In between, there are signal conditioners or processors to feed an improved response in both time and frequency domain. This can be in any direction of the servo, either on the input or sensing the output.

• The signals are conditioned by design to result in minimal error with optimal; dynamic and steady-state response by many analog methods of analysis. (Step,ramp, impulse, Fading etc.).

• They often include more than 1 sensor to match the same rate in time.

• e.g. Position, velocity, acceleration, with incremental or absolute signals.
• there must be a way to measure error with a signal for each spec. of importance.

• it must be precise yet a cost-effective way without adding time or frequency noise or significant delay or loss in phase margin.

• signal conditioners may include amplifiers, filters , f to V, V to f, linear to log, anti-log, power series, parametric, . Etc,

How is Feedforward different than Feedback?

• Negative Feedback (NFB) signals may be processed to resemble what was controlled directly for control by error correction. or enhanced to improve some error measurement.
• Feedforward (FF) signals alter the control of each input parameter** by some knowledge-based algorithm or analog method , rather than just alter the error signal.

• They predict what is needed to control the output , rather than rely on feedback error conditions, they control an input parameter directly for control or compensation.
• PWM is not really a FF example but rather defined just as the input, but if this helps, you may consider it as a simple example.

Control         NFB            FF
=====           =======     ========
Time            Reactive    Predictive
Error Loop      Inside      Outside and maybe inside
Calibration     yes         yes

• Calibration Error correction for gain,offset for any transfer function due to NFG, FF or sensor or actuator due to any variable

• e.g.; including V, [‘C], aging, pressure, humidity,