I picked up a NEC 2190Uxp for cheap (I like square-ish screens, don't ask) hoping I'd get away with just some broken capacitors. No dice, the entire thing is using nothing but Nippon Chemicons, all of which are perfectly fine.
The panel is an Samsung LTM213U6-L01, which has six lamps in two groups on the long sides of the panel. The inverter seems to be a custom design and not the Samsung-suggested reference design (FIF2066-01B, which seems to use six transformers), but rather a custom NEC design (J19I014.00).
Here's the specs for the lamps:
Current: 3-8 mArms, 7.5 typ
Voltage: 740 Vrms typ
Frequeny: 40-60 kHz
Startup voltage ~1500 Vrms
The problem is that the backlight turns off after a few seconds, if the brightness is set to something more than ~20 %. At values below 20 % it works indefinitely (>1 hour).
The backlight inverter is quite a work of art. It uses a single OZ960 controller, which is meant for single-output inverters. They put two output stages on it, using separate FETs and transformers. Each transformer then feeds three outputs through an C-(L||R||Lamp) network.
I've attached a rough schematic of the output stages and some bits and pieces of the feedback networks. Unfortunately there are many vias under components so it's somewhat hard to tell what's going on.
Here's what I figured out so far:
The output from the transformers is ~3500 Vpp, while the voltage across the lamps is 2000-2300 Vpp, which seems to be right on (~710-800 Vrms). The output voltage is undistorted. The low-side of the transformer seems to be tied to ground through a pack of Zener diodes and so it sees around 200 Vpp.
Lamp currents (measured on the cold/return side of each lamp using a high-bandwidth AC current clamp) vary a little bit, from around 5 mArms to a little over 8 mArms; the lamps which are pulling the higher voltages also pull the higher currents. Lamp currents don't look distorted.
The overvoltage protection path is fine. It taps the output of each transformer through two series-connected high voltage (3 kV ea.) ceramic caps and divides it down using two parallel ceramic caps in the lower leg. Then one-way rectification using a BAV99 for each output, then it goes directly to the OVP pin of the OZ960. There's a whole bunch of circuitry which sits on the return wires of each lamp. It looks like they're summing the positive half-wave currents and ultimately feed that into the FB pin of the OZ960 to close the current-control loop, but that only accounts for about 1/3 the extra parts. I suspect they're monitoring the current imbalance across lamps, since I can't come up with an explanation for what you'd otherwise do here.
My suspicion is that the inverter is actually working perfectly fine, as do the protection circuits. I suspect the lamps have just aged enough that the current-imbalance-monitor I'm imagining in there is noticing the uneven lamp currents that I also measured and shuts it down. Since replacing the lamps doesn't really make much sense, I'm probably going to leave it as-is.
I attached two pics of the board for the curious.
Edit: I found the OZ960 datasheet not very clear.
Here's my understanding of it's protection systems:
Ignition mode is enabled when the compensation pin goes above ~2.75 V, because there is no lamp current, the error amp is going to saturate at the positive rail (~3.x V). This raises the output frequency to boost the output voltage. At the same time, the capacitor connected to SST is charged up. The HF duty cycles is proportional to the SST pin voltage in this mode, so effectively the ramp generated on the SST pin has the output amplitude ramp up as well. The OVP pin isn't actually an over-voltage protection pin per-se. It just switches the SST current source off, when 2 V is reached on the OVP pin (so further amplitude ramp-up is prevented), and switches the "CTIMR" current source on. This generates a short ramp using the external capacitor on the CTIMR pin. When that ramp reaches 3 V, the controller turns off and stays latched until a power cycle occurs.
The "OVP goes above 2 V, initiate CTIMR ramp and latch off" part seems to apply regardless of operation mode, i.e. even when the current control loop is operating in the linear region and the error amp output is around 2 V.