A board that is stable on a bench supply can behave very differently on a panel shared with a variable frequency drive. Industrial power is not clean, and the fixes are architectural rather than incremental.
Assume transients, not just noise
Contactors and motor starters generate fast transients every time they operate. These are short, high energy events that bypass ordinary filtering. Transient suppression at the input, sized for the environment rather than the nominal voltage, is the first line of defence.
Isolate the domains
Separate the machine-facing side from the logic side with galvanic isolation, and do it at the interface rather than deep in the design. This bounds the damage from a fault on the machine side and removes an entire class of ground loop problems that are almost impossible to debug on site.
Grounding is a design, not a wire
Decide where your ground reference is and keep return paths short and intentional. Shields should be grounded at one end so they drain rather than conduct. Two well-meaning ground connections at opposite ends of a long run create a loop that will inject exactly the interference you were trying to exclude.
Give the analogue front end its own supply
Sensor measurement is only as clean as the rail behind it. Feeding a high resolution converter from the same rail as a switching load reduces your effective resolution to the noise floor of that switcher. Separate the rails, or at minimum filter locally and reference carefully.
Test with the disturbance present
Validate with a real inductive load switching nearby, not on a quiet bench. It is the cheapest possible way to discover a susceptibility problem while it is still a schematic change rather than a field failure.