Error Correlation
Statistical relationships between successive measurement errors indicate that a model has failed to account for some part of the underlying physical process. If the residual autocorrelation is high, the errors are not random white noise. This suggests that the filter is leaving useful information on the table.
Diagnostic Tool
Engineers use the autocorrelation function to check if the residuals are independent and identically distributed. A non-zero residual autocorrelation at low lags often points to an unmodeled time constant or a sensor bias. It is a sign that the filter needs better tuning.
Impact on Accuracy
When residual autocorrelation exists, the filter’s estimate of its own uncertainty is usually over-optimistic. This can lead to the system ignoring new data even when the model is drifting. In autonomous driving, this might result in a vehicle slowly veering out of its lane.
The presence of these patterns allows developers to identify specific frequencies where the sensor noise is not white. By adding states to the filter, the residual autocorrelation can be reduced. This process continues until the residuals resemble pure noise.
Verification Standard
Passing an autocorrelation test is a standard requirement for the certification of navigation software. Minimizing residual autocorrelation ensures the highest possible reliability for the state estimate.
