Hey there! As someone who's been in the trenches with legacy PLC systems, I can tell you that the glossy technical papers often miss the gritty reality we face daily. Here are the unspoken challenges that keep field engineers up at night:

First, the data extraction nightmare. Those old PLCs speak ancient protocols like DH+, Modbus RTU, or vendor-specific serial communications that modern systems can't understand. You'll spend weeks just trying to get data out, often dealing with undocumented protocols that the original engineers took to their graves.

Then there's the 'data quality black hole' - legacy systems weren't designed for AI. You'll find inconsistent sensor readings, missing historical data, and tags that were never properly documented. The AI models need clean data, but you're feeding them decades of industrial chaos.

Network integration is another silent killer. Trying to connect 20-year-old PLCs to modern Ethernet networks without disrupting production? Good luck. You're often working with serial interfaces that need custom bridging hardware, and any mistake can bring the whole line down.

The human factor is huge too. Your maintenance team knows these systems inside out, but they don't speak 'AI'. Meanwhile, the AI experts don't understand why that particular PLC 5 from 1998 behaves differently on rainy days. Bridging this knowledge gap is harder than any technical integration.

Finally, there's the 'production-first' reality. Technical papers assume you can take systems offline for testing. In the real world, you're implementing AI predictive maintenance while keeping production running 24/7. Every change feels like performing open-heart surgery on a marathon runner.

These are the daily battles that don't make it into the research papers but are the real measure of success in the field!