Weird I2C Problems

You have 4k7 pull-up. You are using 400 kHz frequency. You have a large system with mind-boggling amount of 128 chips and wiring between them which equals a humongous amount of capacitance on bus.

It is a way too optimistic expectation for it to work at all.

You need to do one or multiple of the following to make it work:

1. Slow down the clock. It can't work with that amount of chips and that pull-up resistor. The rise time is too slow for the high time to be long enough.

2. Change to a stronger pull-up resistor to make bus edges faster. It may still require slowing down the clock and reducing the capacitance, as technically 3mA is the limit unless you are using chips that allow for more current.

3. Reduce the bus capacitance. Even if 1 chip equals 5 pF, 128 devices is way beyond the standard operating parameters of 400pF. Also some of the 128 possible addresses are reserved for other purposes like 10-bit addressing, so you really should not use all 128 addresses for 128 slaves, many microcontrollers would even fail to communicate correctly if their I2C peripheral is standard-compliant and adheres to the strict rules of 10-bit addressing. And to be compliant with the specs of Fast Mode, you can't even reach 400 kHz with passive pull-up resistors and the allowed 400pF of capacitance as it will violate the rise time or at least the clock high time requirement. You may need to segment the bus to have less devices per bus segment.

In addition to my comment, I'm going to assume your I2C ground is common with your relay (power) ground. I'm making this assumption because you are able to turn the relays on but not off. What's happening is, when the relays are off, there is negligible current in the ground (return). So the I2C bus functions as it should. After you get the relays on, you now have appreciable current flowing in the ground, which is also the I2C bus return; effectively a ground loop. So this relay current causes voltage drop in the ground wire raising the potential. This is bringing the I2C bus voltages out of spec so the marks and spaces can't be resolved by the slave devices while the relays are on. What you should do to fix this, is give the relay coils independent return paths to keep the I2C bus return "clean". Tie the relay coil ground and bus ground at a single point back to the source.

^{simulate this circuit – Schematic created using CircuitLab}

I agree with the capacitance diagnosis. There are two devices you should be familiar with if you want to stick with I2C: active pull-ups and bus extenders. Active pull-ups (also known as accelerators) detect a rising transition and turn on a current source instead of waiting for a resistor to charge the bus capacitance, this can give you acceptable rise times up to several hundred pF or even a few nF.

Bus extenders (also known as redrivers) split the bus up into two separately-driven domains, chopping up the capacitance into more manageable bits. They can be cascaded and some have active current sources built in.

I have used both of these devices in a long-run configuration where I2C was required by the components we were interfacing and can attest to their effectiveness.

The problem was solved the PCF8575 Boards all had internal 10K pullup resistors, Meaning each time we used a PCF8575 Board it ended up adding more and more resistance, to solve this problem we scratched out a very specific trace