Software filtering
Input errors can be controlled by software filtering (debouncing) of inputs. There are many tried and true debouncing methods. The key is to choose one that best balances the required responsiveness to the input versus the need to tolerate glitches on the input.

For instance, if the system requirements state that you must recognize a button press within 50ms, a good filtering method might be to sample the input in a 10ms ISR and require the input to be seen at the opposite state for at least 40ms (4 consecutive samples) for the debounced state to logically change.

Also, some DSPs, such as TI's TMS320F28XX family, allow you to configure filtering of inputs. The GPIO can have a sampling rate and a qualification count configured. The two together form the sampling window.

For the input qualifier to detect a change in the input, the level of the signal must be stable for the duration of the sampling window width or longer. This is filtering that consumes no instruction cycles after it has been configured, so be sure to use it when you can!

Software can also be used to digitally filter an analog input signal. This is most efficiently accomplished on a digital signal processor, but it can also be implemented on a standard microcontroller too. A FIR (finite impulse response) or IIR (infinite impulse response) filter can be implemented to clean up an input signal that is dirtied by EMI.

Digital filters can provide remarkable performance compared to their analog counterparts. For implementation on a standard microcontroller, there are many 'C' source implementations of these filters freely available on the internet. For implementation on a DSP, almost all DSP vendors provide optimized assembly implementations of these filters that make use of the hardware features of the DSP.

Figure 1. ECG with 60Hz Power Line Interference

For example, consider the case of a medical device that monitors ECG through an A/D converter to detect the R-Wave in a QRS complex. The QRS complex energy is typically in the band from 10Hz to 40Hz. An FIR band pass filter in this band can be extremely effective at eliminating 50/60Hz power line noise as well as other higher frequency noise sources.

The attenuation below 10Hz is more to eliminate uninteresting ECG components, but also eliminates baseline wander and low frequency noise sources. Figure 1 above shows a representative ECG signal with power line interference. Figure 2 below shows the same signal after being filtered by a 100 tap FIR filter with a 0 - 40Hz pass band.

If a transient EMI event corrupts the value of an output port, it is possible that the effect may be minimized by a periodic refresh of the output port values. Of course, in many cases having the output corrupted for any amount of time is unacceptable, but in the cases where it is manageable (an LED output for example), this could be an acceptable solution.

Figure 2. ECG through 100 Tap FIR Filter

This solution can also be applied to configuration registers of on-chip peripherals. A periodic refresh of the configuration of those peripherals might make the temporary corruption of one of those registers tolerable. However, sometimes writing to a peripheral configuration register can cause a reset or other disruption in the operation of the peripheral, so use care when employing this method.

In many cases, the designer has a choice between using a level-triggered or an edge-triggered interrupt. If it is feasible, the level-triggered interrupt should be chosen. Whereas an edge-triggered interrupt typically does not impose any minimum requirement on the interrupt event pulse width, a level-triggered interrupt will.

Typically, an interrupt controller samples the interrupt inputs at some defined frequency (once per instruction cycle, for instance). A level-triggered interrupt will require the event to be present two samples in a row before an interrupt is generated. The minimum width is the sample period. An edge triggered interrupt will look for two consecutive samples that indicate the intended transition.

Since the transition can occur almost immediately before the next sample, there really is no minimum width imposed on the interrupt event. The net effect is that noise is more likely to trigger an edge triggered interrupt than a level triggered interrupt. Microcontrollers sample interrupts using different methods. Understanding how your microcontroller samples interrupts is the key to determining which type of interrupt is less sensitive to noise than the other.