Identify the SQL query causing spikes

One way would be to set up an Extended Events session that captures statements that consumes more than a certain amount of CPU time. Given that there is some red tape with saving the data to blob storage, it is easier to use a ring buffer as a target, and that should work well under the assumption the issue is due to some occasional wild query. For more information to do that, see https://zcusa.951200.xyz/en-us/azure/azure-sql/database/xevent-code-ring-buffer?view=azuresql&tabs=sqldb

However, it could also be matter of an application that submits a lots of shorter statements with hard-coded values, which causes a lot of compilation. This will not be detected by the X-event session I discussed, and overall it can be more difficult to track down. But start with the XE session as I suggested.