In a variety of fields, system inversion is often required in order to determine inputs from measured or for desired outputs. However, inverse systems are often non-proper in the sense that they require differentiators in their realisation. This leads to numerical difficulties associated with the computer implementation of their mathematical models. To overcome these problems, approximate inversion also referred to as filtered inversion is proposed for systems modelled by bond graphs. Generic configurations of right and left filtered inverse bond graph models are proposed with dynamic structural conditions on the filters so that the resulting composite bond graph represents a proper system suitable for effective numerical implementations.