How do I maintain high INSERT-performance on PostgreSQL

I’m working on a project which is parsing data from measurement files into a Posgres 9.3.5 database.

At the core is a table (partitioned by month) which contains a row for each measurement point:

CREATE TABLE "tblReadings2013-10-01"
(
-- Inherited from table "tblReadings_master":  "sessionID" integer NOT NULL,
-- Inherited from table "tblReadings_master":  "fieldSerialID" integer NOT NULL,
-- Inherited from table "tblReadings_master":  "timeStamp" timestamp without time zone NOT NULL,
-- Inherited from table "tblReadings_master":  value double precision NOT NULL,
  CONSTRAINT "tblReadings2013-10-01_readingPK" PRIMARY KEY ("sessionID", "fieldSerialID", "timeStamp"),
  CONSTRAINT "tblReadings2013-10-01_fieldSerialFK" FOREIGN KEY ("fieldSerialID")
      REFERENCES "tblFields" ("fieldSerial") MATCH SIMPLE
      ON UPDATE CASCADE ON DELETE RESTRICT,
  CONSTRAINT "tblReadings2013-10-01_sessionFK" FOREIGN KEY ("sessionID")
  REFERENCES "tblSessions" ("sessionID") MATCH SIMPLE
  ON UPDATE CASCADE ON DELETE RESTRICT,
  CONSTRAINT "tblReadings2013-10-01_timeStamp_check" CHECK ("timeStamp" >= '2013-10-01 00:00:00'::timestamp without time zone AND "timeStamp" < '2013-11-01 00:00:00'::timestamp without time zone)
)

We are in the process of populating the table with data that has already been collected. Each file represents a transaction of around 48,000 points and there are several thousand files. They are imported using an INSERT INTO "tblReadings_master" VALUES (?,?,?,?);

Initially the files import at a rate of 1000+ inserts/sec but after a while (an inconsistent amount but never longer than 30mins or so) this rate plummets to 10-40 inserts/sec and the Postgres process rails a CPU. The only way to recover the original rates is to perform a full vacuum and analyze. This is ultimately going to be storing around 1,000,000,000 rows per monthly table so the vacuum takes some time.

EDIT: Here is an example where it ran for some time on smaller files, and then after larger files started it failed. The larger files look more erratic but I think it is because the transaction is only commited at the end of a file, around 40sec.

There will be a web front end selecting some items but no update or deletes and this is seen with no other active connections.

My questions are:

1. How can we tell what is causing the slowdown/rail the CPU (this is on Windows)?
2. What can we do to maintain the original performance?