Whisperer a failing bearing in motor current

I used to work at Pak-American Fertilizers in Daudkhel, Mianwali, a urea and ammonia plant that makes its own power. Every week, multiple foremen walked the plant, checked the motors, and wrote a report; if anything sounded bad, a supervisor came to check it. The care was for a reason: if a motor causes a problem and the plant stops, starting it again can take more than a week, and every lost day is very expensive. So the most important places got backup motors with sensors on their bearings. Most motors had no sensor at all. But every motor on a drive already has its current measured, for free. Can that current tell a damaged bearing from a healthy one? Papers on the public Paderborn bearing data often report about 99%. We found that most of that comes from letting the model see the test bearings during training. Our same model scores 0.95 that way and 0.81 when tested honestly on bearings it has never seen (balanced accuracy; 0.50 is a coin flip). Bearing Whisperer is a web app for the maintenance technician who must decide which motor to check first: - Plant: 32 motors on one screen, colour-coded "check bearing", "ask a person", or "healthy". - Motor: where the detector puts this motor, in words (never a fake percentage), and its "hiss picture" - Plan my shift: say how much time you have. The AMD assistant calls the plant's tools and returns an ordered visit list with draft work orders that a person approves. - Explain this chart: a vision model on AMD reads the exact chart the trainee sees and explains it. - Proof: every test round, the failures included, as charts. All language and vision work runs on open models served on AMD GPUs through AMD Token Factory: Gemma-4-31B-IT, with Qwen3.8-27B as backup. They write the technician notes, answer follow-up questions, plan the shift with native tool calls, and read charts from images. The detector itself is a small random forest that runs on any CPU.