YUANZHI ROBOTICS 機械人聲學成像載荷指南：檢測洩漏與局部放電

By the engineering team at YUANZHI ROBOTICS. We design and build sensor payloads for robot dogs, UGVs and drones. Originally published on our site; AI-assisted draft, reviewed by our engineers.

Thermal cameras find faults that are already hot. Acoustic imaging finds two kinds of fault before they are: gas and compressed-air leaks, and partial discharge on high-voltage equipment. Both make ultrasound, both are invisible, and both are easy for a robot to pick up on a routine round if the payload is built properly. This guide covers what the sensor is, where it is useful, how we mount and power it on a quadruped or UGV, and how the readings become something an inspector can act on.
Inspection interface on the robot: every reading, acoustic ones included, is attached to a named inspection point and compared with the threshold set for that point

What an acoustic imager is
An acoustic imaging camera is an array of small MEMS microphones, usually 64 to 128 of them arranged in a spiral or ring on a flat plate, with a visible camera in the centre. A processor runs beamforming on the microphone signals and produces a map of where sound is coming from, which is drawn on top of the camera image as a coloured spot. The useful band is ultrasonic, from about 20 kHz up to 50 or 60 kHz, where leaks and electrical discharge are loud and ordinary plant noise is quiet.

Two outputs matter for inspection. The first is the localized source: a position in the image and a level in dB, so the operator sees that the noise comes from the third flange from the left, not the pipe in general. The second is the pattern of the sound over the AC cycle, called a PRPD pattern (phase-resolved partial discharge). Corona, surface discharge and floating-component discharge produce different patterns, and the software classifies them.

What it finds, and what it cannot
Compressed-air leaks in plants, where each leak costs energy around the clock. The imager quantifies the leak rate from the ultrasonic level and distance, which makes a leak survey a ranked repair list.
Gas leaks at flanges, valve stems and seals, including SF6 on open-air GIS flanges and hydrogen on generator seals. A pressurized leak is detected by its sound before a gas sensor could sample it, and from a distance.
Partial discharge on open-air equipment: insulator strings, bushings, cable terminations, surge arresters, open switchgear. Corona on a contaminated insulator, surface tracking, and floating hardware are the typical finds.
Mechanical noise: bearings, belts and cavitation produce broadband ultrasound that shows up as a source on a motor or pump.
What it cannot do is look inside a sealed metal enclosure. Metal-clad switchgear and GIS compartments need a transient earth voltage (TEV) sensor on the panel surface or a UHF sensor; we combine these with the acoustic imager on substation payloads, one for open-air equipment and the other for the switchgear room.

Mounting on a robot dog or UGV
The imager wants a clear line of sight and a steady platform. On our payloads it sits on the two-axis pan-tilt head next to the visible and thermal cameras, so the three see the same scene and the acoustic overlay can be checked against the thermal image at the same inspection point. A robot stops at each point for a few seconds; the head turns to the stored pan and tilt angles, the imager integrates for a second or two, and the result is stored. Quadrupeds move the body slightly while standing, which does not matter at these integration times.

3D render of a Unitree B2 inspection payload: the sensor head at the front carries the cameras and the acoustic array on a common pan-tilt unit Details that decide whether the mount works:

keep the array face free of anything that reflects ultrasound within about 30 cm, including the robot's own antennas and the thermal camera hood;
mount the array plate on rubber isolators, because the robot's leg motors and the payload fans produce ultrasound of their own;
route the Ethernet and power through the pan-tilt slip ring or a service loop with enough slack for the full tilt range;
give the array a rain cover that leaves the microphone holes open, since a sealed window attenuates the band you need.
Weight and power are modest: the imager module is 0.5 to 1.5 kg and draws 5 to 15 W at 12 or 24 V, and most of that is the processor. A complete head with visible, thermal and acoustic sensors is 2 to 3 kg, within the payload budget of a Unitree B2, DEEP Robotics X30 or any inspection UGV. On a Unitree Go2 EDU the head is too heavy; there a lighter 64-microphone module without its own gimbal can be fixed to the body.

From a coloured spot to an inspection result
A single acoustic image is a picture; an inspection needs numbers that can be compared between rounds. The payload software does four things with each measurement:

stores the peak level in dB, the source position in the image and the PRPD classification together with the robot pose, the head angles and the time;
attaches the measurement to the named inspection point, with the threshold set for that point (a bushing and a compressor hose do not share a limit);
trends the level at that point over the last rounds, because a rising level at the same place is the strongest sign of a developing fault;
records the weather from the payload's environmental sensor, since rain and wind change the noise floor and explain most of the variation between rounds.
Alarms are graded: a new source above threshold is a warning, a rising trend over three or more rounds is an alarm, and a confirmed PRPD pattern of surface discharge is escalated. Each is pushed to the operator's reporting system with the image, the thermal image from the same point and the recommended follow-up, typically an offline test or a closer look with a handheld instrument.

Acceptance testing
We test acoustic payloads with two reference sources: a small ultrasonic emitter at a known distance for alignment and level, and a partial-discharge simulator for the PRPD classifier. Acceptance is passed when the robot localizes the emitter within a set pixel distance at 5 m on three consecutive rounds and classifies the simulator pattern correctly. The same two sources are used on site after commissioning and whenever the head is remounted.

The same sensor head design is reused across several robot platforms; only the mounting plate and the cable set change

When to add it
Add an acoustic imager when the site has pressurized systems or open-air high-voltage equipment: substations, power plants, compressor stations, chemical plants, large factories with compressed-air networks. For indoor switchgear only, a TEV probe is cheaper and more useful. For sites where the question is mostly temperature, a thermal gimbal alone is the right first payload and the imager can be added later on the same head.

If you are specifying an inspection payload and want to know whether an acoustic imager belongs on it, send us the site type and the equipment list; we will tell you what we would mount and why.

FAQ
What does an acoustic imaging camera detect on a robot?

Anything that makes ultrasound: gas and compressed-air leaks (turbulent flow at 20 to 60 kHz), partial discharge on insulators, bushings, cable terminations and open switchgear (corona, surface and floating discharge), and mechanical faults such as bearing noise. It does not see inside sealed metal enclosures; for metal-clad switchgear and GIS a TEV or UHF probe is used instead.

How far away can the robot be?

Leak detection works at 1 to 10 m depending on leak rate and background noise; partial discharge on open-air equipment is typically localized at 3 to 15 m, which keeps the robot outside the safety clearance of live equipment.

Does the array need to be calibrated?

The array itself is factory-calibrated and stable; what needs checking is the acoustic-to-camera alignment after any remount, which is done with a small ultrasonic reference source placed in front of the robot during factory acceptance.

Can it work in a noisy plant?

Yes, because the processing filters to the ultrasonic band (above about 20 kHz), where machinery and wind contribute little. Rain and strong wind still raise the noise floor, so the software records the weather with each reading and the threshold is set per inspection point.

How heavy is the payload?

The imager alone is 0.5 to 1.5 kg and 5 to 15 W. Mounted on a two-axis head with a visible and thermal camera the complete head is 2 to 3 kg, which fits a Unitree B2, DEEP Robotics X30 or a wheeled UGV without changing anything else on the robot.

More guides: Robot payload guides. Questions about a specific robot or site: lizhongpan@yuanzhirobot.com