Short answer
Gauge reading data is a set of images of dials and meters with a bounding region, scale range, unit, needle position and the true reading for each. Facilities, plants and utilities hold the raw material in round sheets and inspection photos, but readings logged by operators are the only labels many sites have. General object datasets do not teach dial reading; a purpose-labeled set does.
The buyer's problem
Plenty of buildings and plants still depend on analog instruments: pressure gauges on pipes, temperature dials on boilers, panel meters on switchgear, sight glasses on tanks. Replacing them with sensors is a capital project, so a camera-equipped system that reads them is an attractive retrofit. It is also a task a mobile robot can perform on a round: stop in front of a gauge, capture an image, report a value.
Detecting that a gauge exists is the easy part. Reading it is a measurement problem. The needle sits at an angle between ticks; the scale may be logarithmic or have several rings; the unit and range differ per device; glass reflects light; and the camera is rarely perpendicular. General detection sets are built around object labels. Microsoft COCO, for example, provides about 2.5 million labeled instances across 91 object categories in 328,000 images (Lin et al., 2014). That coverage helps find objects in a scene, but its labels name the object and outline it; they do not say what number the object displays. A reading model needs its own annotation layer.
This page covers where to source such images and how to structure them. For other uses of the same instruments in hands-on tasks, see measuring instrument use. The maintenance hub and the humanoid data guide frame the wider cluster.
What counts as a gauge reading record
| Field | Description | Note |
|---|---|---|
| Image | Full frame as captured | Keep the original, not only a crop |
| Gauge region | Box or polygon around the dial face | Needed for detection and cropping |
| Instrument type | Pressure, temperature, flow, voltage and so on | Drives scale expectations |
| Scale definition | Min, max, unit, linear or nonlinear, number of rings | Often missing from raw photos |
| Needle or pointer geometry | Pivot and tip points, or angle | Enables geometric reading |
| True reading | Value as read by a person and checked | Label for training and evaluation |
| Condition bands | Green, amber, red zones | Link to alarm limits |
| Capture conditions | Distance, angle, lighting, glare, fog, dirt | Basis for robustness analysis |
The true reading is the expensive field. Two readers can disagree by a tick or more on an analog dial, so a dataset that records a single reader's value has an unstated error.
Supplier types and what they hold
Facilities and property operators. Their operator rounds sheets list gauge tag, value and time. Some also attach photos. The operator rounds and walkdowns page describes how to turn a round into a labeled sequence.
Plants and process operations. They hold control-room logs that can act as truth values when a sensor reading is available for the same moment as the photo. This gives an independent label without human reading.
Utilities and field service organizations. Technicians photograph meters and gauges to document visits, often with a phone. These images vary in angle and lighting, which is useful for robustness but demands careful quality control.
Inspection departments. Photos attached to inspection reports are tied to dates and equipment IDs; the inspection photos catalog entry describes the nearby asset class.
Existing images or new capture
Existing images arrive with whatever labels the site produced: tag number, logged value, maybe a comment. They represent real conditions, including bad ones. They are rarely sufficient alone, because the logged value may have been rounded, and photos taken as evidence are framed around the person's needs rather than the scale.
A new capture can fix distance, angle ranges, lighting conditions and a second reader per image. It needs site access and a safety plan for areas with hazards, discussed in recording in hazardous spaces. Some teams combine both: existing photos for diversity of conditions, a small purpose-captured set for clean evaluation.
What the data can support, and what it cannot
Labeled gauge images can support detection, crop-and-read models, scale recognition and out-of-range checks. Paired with a logged sensor value, they support evaluation of reading accuracy under glare, angle and soiling.
They cannot tell a robot how to reach the gauge, open a cabinet, clean a lens or avoid a hazard. They do not teach the meaning of a reading, such as whether a pressure is normal for the process. They do not show whether the reading was acted on. And they do not prove a model is safe to automate round-based decisions; that requires a site-specific validation. Object recognition data and manipulation data are separate assets, as described in object recognition versus manipulation data.
For defect-type annotation conventions that parallel this work, see defect image datasets.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: gauge_reading_images_v0
source: facility_round_photos from 3 sites (site_011, site_017, site_023)
images:
count_range: set by buyer's sampling plan
format: jpeg, original resolution retained
annotations:
gauge_polygon: per image
instrument_type: pressure | temperature | flow | panel_meter
scale: {min: 0, max: 160, unit: psi, mapping: linear}
pointer: {pivot_xy: [412, 388], tip_xy: [455, 321]}
reading: {value: 62.5, reader: reader_A, second_reader: reader_B}
condition_band: green
capture_conditions: {angle_deg_off_axis: 25, glare: moderate, lens: phone_wide}
independent_value: control_log_value at same minute where available
split: held out by site_id
rights: AI-training use of images permitted; redaction of tag plates and signage as agreed
From reading values to reading context
A value alone rarely drives a decision. The same needle position can be normal on one pump and an alarm on another, so a useful package ties each gauge to its asset record, its normal range and its alarm limits. Panels with fault lamps or digital displays add a second reading task, covered in error codes and HMI screens, and many sites mix both on the same equipment.
If you plan a round-based robot task, ask whether the supplier can provide sequences: the same gauge photographed on consecutive rounds, with readings, so trends and sensor drift can be studied. Single photos from many gauges and long sequences from a few gauges answer different questions, and the sampling plan should say which you are buying.
Quality and acceptance criteria
- Two-reader agreement reported, with the tolerance expressed in scale units rather than as a percentage of the image.
- Scale metadata present on each gauge type, not inferred from the image.
- Evaluation images from sites not used in training.
- Distribution of off-axis angles, glare and dirt reported so you can see what the model has not seen.
- Cropping rule documented so a model trained on crops can be tested on full frames.
Permissions and provenance
Gauge photos often show more than gauges: nameplates, panels, adjacent equipment, people and process layout. A plant operator may treat the arrangement as confidential, and utilities may have security considerations about critical infrastructure. Diligence questions include whether the supplier owns the photographs, whether employees or contractors took them, whether third-party vendor equipment is shown, and what the supplier's site agreement allows. Redaction of signage and tags helps with some of these concerns but does not settle rights.
Questions to ask suppliers
- How were the images captured, and by whom, with which devices?
- What is the source of each reading label, and was a second reader used?
- Is the scale range and unit recorded per gauge, or only the logged value?
- Do photos have corresponding sensor or control-system values?
- Which gauge types and conditions (glare, dirt, low light) are represented?
- How many sites and distinct gauge models are included?
- What can appear in the background, and what redaction is applied?
- Who can authorize AI-training use of the images?
Explore sourcing options
If you need labeled instrument images, describe the gauge types, label fields and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources. A request does not guarantee a match. For adjacent materials see inspection reports and preventive maintenance checklists, whose entries often record the readings that serve as labels.
Explore sourcing and licensing options with SourceX
Sources
- Microsoft COCO: Common Objects in Context — object-level labels in a general-purpose detection dataset, used here as a contrast with reading labels.