make it impossible to do more than guess at this but obviously the sensor is reading either a vibration signal or variations in the belt tension as it operates.
My approach would be:

First:
Determine where the error comes from (vibration, sensor mounting position, ect).
Second:
Correct the mechanical causes to minimize the error signals
Third:
Tare out the natural input to the sensor that is the real weight of the conveyor. (Zero the scale with the machine stopped)
Forth:
Average the input signals over sufficient time that the empty belt just gives a steady 0 on the output.

The above steps will give you the best response time for this undefined system. If the response time is too long the basic engineering behind this probably needs to be changed to improve the resonse time.
More details would be nice as they would lead to approaches that are governed by the posted data.

Good luck Terry