As you probably already know a TLE is a strange animal. It does not really contain proper orbital elements, but instead is engineered with one purpose; to be fed into SGP4 so that that will generate reasonable position information for at least a few days around the TLE's epoch. See @Tristan's answer to Is SGP4 propagation necessarily more accurate near the epoch chosen for TLE generation? and answers to How to obtain UTC of the epoch time in a satellite TLE (two line element)? and How does SGP4 work? for more on that.

SGP4 includes approximations for several effects beyond Keplerian orbits, including a modest "lumpy gravity" model for Earth, continuous atmospheric drag, and gravitational perturbations from the Sun and Moon. For more on that last one see answers to SGP 4 for Geostationary Satellite and How do SDP4's "Deep space" corrections to SGP4 account for the Sun's and Moon's gravity? and SGP4 on Systems Tool Kit (STK); how to check if SDP4 deep space correction is implemented? and maybe Differences between SGP8 and the standard SGP4? Is it ever used in practice?

From these I would like to convince you that a direct interpretation of the numerical values in the TLEs should be taken with several grains of salt. Large, sudden changes in mean anomaly (especially to higher orbits) or in inclination from one TLE to the next might indeed indicate a propulsive maneuver, but smaller changes could indicate a change in atmospheric temperature and density in LEO due to solar activity or even a mixture of noisy individual position and velocity measurements from satellite tracking (radar, visual/telescope).

So I think it will be very difficult to flag a propulsive maneuver with certainty and distinguish it from day-to-day noise in TLEs unless it is large or unless one group of TLEs are inconsistent with a following group of TLEs.

To that end, what you can do is for each TLE use SGP4 to predict a position and velocity at the epoch dates and times of several TLEs before and after it, say within a few day or week-long interval around the given TLE's epoch. If you see a clear inconsistency over several TLEs that might warrant further investigation.

For more background before you start that, see

• Can the TLE epoch be listed as in the future? How?
• How to obtain UTC of the epoch time in a satellite TLE (two line element)?
• Time, UTC, Julian Date, TLE epoch - how are they related quantitatively?
• What does it mean to ask for a date beyond epoch time?
• How to find T₀ and other parameters from a TLE to calculate an approximate mean anomaly?
• Is the amount of accurate information in a TLE equal to that of a single state vector?
• How frequent do TLEs of LEO satellites change?

For Python and TLE propagation using SGP4 one very handy option is https://rhodesmill.org/skyfield/

As you probably already know a TLE is a strange animal. It does not really contain proper orbital elements, but instead is engineered with one purpose; to be fed into SGP4 so that that will generate reasonable position information for at least a few days around the TLE's epoch. See @Tristan's answer to Is SGP4 propagation necessarily more accurate near the epoch chosen for TLE generation? and answers to How to obtain UTC of the epoch time in a satellite TLE (two line element)? and How does SGP4 work? for more on that.

SGP4 includes approximations for several effects beyond Keplerian orbits, including a modest "lumpy gravity" model for Earth, continuous atmospheric drag, and gravitational perturbations from the Sun and Moon. For more on that last one see answers to SGP 4 for Geostationary Satellite and How do SDP4's "Deep space" corrections to SGP4 account for the Sun's and Moon's gravity? and SGP4 on Systems Tool Kit (STK); how to check if SDP4 deep space correction is implemented? and maybe Differences between SGP8 and the standard SGP4? Is it ever used in practice?

From these I would like to convince you that a direct interpretation of the numerical values in the TLEs should be taken with several grains of salt. Large, sudden changes in mean anomaly (especially to higher orbits) or in inclination from one TLE to the next might indeed indicate a propulsive maneuver, but smaller changes could indicate a change in atmospheric temperature and density in LEO due to solar activity or even a mixture of noisy individual position and velocity measurements from satellite tracking (radar, visual/telescope).

So I think it will be very difficult to flag a propulsive maneuver with certainty and distinguish it from day-to-day noise in TLEs unless it is large or unless one group of TLEs are inconsistent with a following group of TLEs.

To that end, what you can do is for each TLE use SGP4 to predict a position and velocity at the epoch dates and times of several TLEs before and after it, say within a few day or week-long interval around the given TLE's epoch. If you see a clear inconsistency over several TLEs that might warrant further investigation.

For more background before you start that, see

• Can the TLE epoch be listed as in the future? How?
• How to obtain UTC of the epoch time in a satellite TLE (two line element)?
• Time, UTC, Julian Date, TLE epoch - how are they related quantitatively?
• What does it mean to ask for a date beyond epoch time?
• How to find T₀ and other parameters from a TLE to calculate an approximate mean anomaly?
• Is the amount of accurate information in a TLE equal to that of a single state vector?
• How frequent do TLEs of LEO satellites change?