Rank the usual causes by frequency: underestimated costs and slippage, look-ahead-bias in the research code, a signal calculated differently live than in research, overfitting to the test sample, and a genuine market-regime change. The first three are bugs and the last two are statistics.
Diagnose it by comparing trade by trade, not by comparing returns. Run the backtest over the live period and match fills: same signals at the same times means the divergence is execution; different signals means the implementation differs from the research.
Expect the honest version of degradation even with no bugs at all. A strategy selected as the best of many will regress toward the mean, so budget for roughly half the backtested Sharpe and treat anything better as a pleasant surprise.
Related: incubation, strategy-monitoring, transaction-cost-modelling, overfitting