Follow one reading from dial to user
Reading begins with an image showing the scale and needle clearly. Locate the correct instrument, then identify geometric points such as center and needle tip before converting the angle using its range and unit. Correct detection does not establish correctness of every later stage.
Confirm the actual scale, including direction, sweep and nonlinear ranges. An incorrect configuration may produce smoothly changing but wrong numbers. Digital displays require decimal points, minus signs and units, not partial text recognition.
Associate the result with instrument identity, capture time and validity before presentation or storage. Stage timestamps distinguish processing delay from delivery delay. Quality states prevent old or unclear observations appearing as successful fresh readings.
Why unreadable is a necessary result
Zero can be genuine when a process stops or pressure falls. Do not use it to represent unreadable images. Reflections, fogging, occlusion and missing dials need distinct states and reasons so maintainers can correct imaging or use a backup inspection method.
Model confidence is an internal signal, not evidence of numerical correctness at that percentage. Validate rejection thresholds using real examples and report both answered errors and rejection rates. Forcing every image to yield a value can conceal dangerous guesses.
Acceptance should compare with people or reference instruments under agreed definitions. Separate image-reading error from instrument uncertainty. A camera does not calibrate a gauge or replace independent protective signals for emergency or safety functions.
Keep readings meaningful during trials and maintenance
Begin in shadow mode alongside the existing method without control write-back. Cover shifts, scale ranges and lighting. This establishes appropriate scope before workflow integration; a few demonstration images cannot validate an actual site.
Recheck geometry and scale conversion after camera, lens, lighting or software changes. Version settings with results and assign ownership of missing data, blur and deteriorating quality.
Review each additional reading point, even when sharing a model: scales and acceptance thresholds may differ. Actual integration, device support and service terms require confirmation with an SI partner rather than inference from the pipeline diagram.