Latest / Internet of Things with Fexingo: Connected Devices, Sensors, and Industrial IoT / How IoT Sensors Are Preventing Tunnel Boring Machine Disasters
Transcript
- Lucas: So last month, a tunnel boring machine under downtown Bangkok lost a cutterhead to what the operator called 'unforeseen ground conditions.' Cost them four months and about 30 million dollars. The thing is — the ground wasn't unforeseen. Sensors on that machine had been flagging abnormal vibration patterns for nearly two hours before the failure. Luna: Wait — the sensors caught it, and nobody acted? Lucas: Not exactly nobody. The data was there, but the control room software was set up to alert only on hard thresholds — temperature above 90 degrees, vibration above 12 millimeters per second. The precursor signals were subtle shifts in frequency, not amplitude. The kind of pattern a human analyst might spot, but the automated system was blind to. Luna: So it's a software gap, not a hardware gap. Lucas: Exactly. And that's the story we're digging into today — how a new generation of sensor fusion and edge computing is turning these multi million dollar moles from blind diggers into machines that can essentially feel the ground ahead. I'm Lucas. Luna: And I'm Luna. This is Internet of Things with Fexingo. And Lucas, you mentioned Bangkok — was that the same tunnel project that made headlines in May? Lucas: That's the one. The Orange Line extension, running east-west under the Chao Phraya River basin. The TBM was a Herrenknecht S-939, about 120 meters long, boring a 6.7-meter diameter tunnel through a mix of clay and sand. The ground conditions there are notoriously variable — you can go from stiff clay to saturated sand lenses within a single ring. Luna: And that's exactly where the sensors should shine. What are we talking about — vibration, temperature, what else? Lucas: Standard modern TBM already carries about 200 to 300 sensors. Thrust cylinders have pressure transducers. The cutterhead has torque meters and temperature probes. There are inclinometers along the shield, strain gauges on the segment erector — the list goes on. But the typical approach is to log these to a database and review them after the shift, or trigger alarms for single-parameter exceedances. What's changing is the real-time fusion of these signals into a composite picture of ground risk. Luna: Give me a concrete example from the Bangkok operation. Lucas: Sure. After the failure, the contractor brought in a specialist firm called Geosense — they've been retrofitting TBMs with an edge computing module they call the 'GroundSight' unit. It ingests all 200-plus sensor streams at 50 hertz, runs a machine learning model trained on past TBM failures, and outputs a single continuous risk score from zero to one hundred. During a test on the remaining drive of the Orange Line, the unit detected a developing anomaly about 15 seconds before the cutterhead torque spiked — enough time to halt rotation and reduce thrust, avoiding a potential face collapse. Luna: Fifteen seconds is not a lot. But in TBM ops, that's a lifetime. Lucas: It really is. When you're spinning a 120-ton cutterhead at 2.5 revolutions per minute, stopping that inertia takes several seconds. And if the ground collapses onto the cutterhead, you're looking at weeks of grouting and hand-mining to free it. Fifteen seconds of early warning can mean the difference between a controlled stop and a million-dollar rescue operation. Luna: What kind of sensor was most valuable in that detection? Lucas: The team told me it was actually a combination of three: the cutterhead vibration accelerometers, the shield strain gauges, and the muck temperature sensors. The accelerometers picked up a shift in the frequency signature — normally the cutterhead produces a dominant frequency at its rotational speed, but they started seeing a second harmonic appear, which indicated the rock was fracturing differently. The strain gauges on the shield showed a slight increase in hoop stress. And the muck temperature, which is measured as the excavated material passes through the screw conveyor, dropped by about 2 degrees Celsius — suggesting water inflow from a sand lens. Luna: So no single sensor would have told the story. It really needed that composite. Lucas: Exactly. And that's the hard part — building a model that can fuse all these signals and distinguish a real precursor from normal operational noise. A TBM changes its vibration profile when the cutterhead wears, when the geology changes, when the operator adjusts the advance rate. The model has to be robust to all those normal variations and only flag the dangerous ones. Luna: How many TBMs are running this kind of edge AI today? Lucas: It's still early. Geosense says they've retrofitted about 14 machines worldwide. The major manufacturers — Herrenknecht, Robbins, NFM — are all developing their own systems. The European Union is actually funding a project called Tunnel-Sense that aims to standardize sensor data formats and share failure case studies across operators. The goal is to have a common benchmark dataset by 2028, so any TBM can be retrofitted with a certified detection model. Luna: That's the kind of standardization that really moves an industry forward. And it's not just about saving money on repairs, right? There's a safety angle. Lucas: Absolutely. Tunnel boring is dangerous work. In 2019, a TBM in India hit a gas pocket and exploded, killing four workers. In 2014, a TBM in Seattle caused a sinkhole that swallowed a truck. The sensors can't prevent every catastrophe, but they can dramatically reduce the odds. Especially when you combine ground sensing with atmospheric monitoring for methane and hydrogen sulfide — which a lot of TBMs already do, but again, often as standalone alarms rather than integrated risk scores. Luna: You mentioned the Bangkok test had a 15-second warning. Has any system actually prevented a full collapse in real operations? Lucas: Yes. In December 2025, a TBM boring a sewage tunnel in Madrid detected a fault zone about 3 meters ahead using seismic reflection sensors mounted on the cutterhead — those are a different class of sensor, essentially active sonar. The system alerted the operator to reduce thrust and increase grout pressure. They ended up passing through a 2-meter-wide cavity that would have caused a surface collapse if they'd hit it at full speed. That's documented in the Tunnel-Sense preliminary report. Luna: Seismic reflection on a TBM — that's essentially a small earthquake source on the machine, right? Lucas: Right. They fire a controlled acoustic pulse into the ground ahead and measure the reflections. It's been used in geotechnical surveys for decades, but putting it on a moving, vibrating TBM and processing it in real time is a serious signal processing challenge. The Madrid machine used a system from a company called InnoSight that runs the processing on an onboard GPU. They claim they can see up to 15 meters ahead with about 20-centimeter resolution. Luna: That's game-changing. If you can see a boulder or a fault zone before you touch it, you can adjust your approach. Lucas: Exactly. And that's where the industry is heading — not just monitoring the machine's health, but actually mapping the ground in real time. It's a shift from reactive to predictive ground control. Luna: Quick honest thing — a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that's literally what funds making this many of these episodes. If today's tech conversation gave you something usable, that's the reason we can keep it ad-free. Lucas: Yeah, it's a small group that makes this possible. And we're grateful. Alright — back to the tunnel. One challenge that keeps coming up is power. A TBM can draw several megawatts, but the sensor and computing gear needs clean, conditioned power. On the Bangkok machine, the GroundSight unit actually has its own battery backup and a UPS because the main power can flicker when the cutterhead hits a hard patch. Luna: That's a classic IoT problem — you can have the best sensors in the world, but if the power dips for 100 milliseconds, you lose data and the model goes blind. Lucas: Exactly. And underground, it's not easy to get a backup generator. So the unit has to be rugged — IP67 rated, vibration isolated, and able to survive a 50-degree Celsius environment. The Madrid system actually uses a passive cooling loop because there's no reliable compressed air for active cooling. Luna: What about data transmission? You can't exactly run fiber to a moving TBM. Lucas: Most TBMs have a trailing cable that carries power and data back to the surface, but it's frequently damaged. So many operators now use a wireless mesh network along the tunnel — nodes every 50 meters that relay data from the TBM to the surface via a series of repeaters. The bandwidth is limited — maybe 10 megabits per second — but enough to stream the sensor summaries and risk scores. The raw high-frequency data is stored locally and retrieved on maintenance shifts. Luna: That mesh network itself is an IoT deployment. Nodes have to be battery-powered or powered from the tunnel lighting circuit. Lucas: Right, and they need to be intrinsically safe if there's any risk of explosive gases. That adds cost. A typical tunnel mesh might run 50 to 100 nodes for a 2-kilometer drive, at about $800 per node. But compared to the cost of a single TBM breakdown — which can easily hit $100,000 per day in lost progress — it pays for itself quickly. Luna: So what's the next frontier? Where do you see this going in the next three to five years? Lucas: I think we'll see fully autonomous TBM operation in certain controlled conditions. Herrenknecht already has a mode they call 'AutoAdvance' that maintains a target advance rate by adjusting thrust and cutterhead speed based on sensor feedback. That's essentially level 2 autonomy. The next step is level 3 — where the machine can detect an anomaly and decide autonomously to stop or change behavior, only alerting the human operator after the fact. That's probably 3 to 5 years away. Luna: And that would require regulatory approval, I imagine. Lucas: Massive. Tunnel boring is heavily regulated, and the liability for collapse is enormous. No contractor wants to be the first to say 'we let the machine decide.' So I think we'll see a hybrid approach for a long time — the machine recommends, the human approves. But the sensors and the models will get better, and the human will approve more and more of the machine's recommendations. Luna: It's a fascinating shift. From blind digging to nearly surgical precision. Lucas: It really is. And it's happening faster than most people realize. The Bangkok failure was a wake-up call — but the technology to prevent it was already in the pipeline. Now it's just a matter of deployment and trust. Luna: Alright, that's our look at tunnel boring sensors. I'm Luna. Lucas: And I'm Lucas. Thanks for listening.