Could smart audio sensors have prevented the train derailment in Ohio?

At around 9:00 p.m. on February 3rd, 2023, a Norfolk Southern Railway train derailed on the main track near East Palestine, Ohio. Eleven tank cars carrying hazardous materials derailed, igniting toxic fires and causing an environmental disaster. At BMT, we believe this derailment could have been prevented if smart audio sensors had been used instead of traditional temperature sensors.  

According to the NTSB report, the train passed by several hot bearing detectors (HBD) that reported an increase in temperature as the train moved past them. The hot bearing detectors were located between 10 and 20 miles apart. Over 30 miles, the train’s bearings went from normal readings to indications of critical high temperatures. Even though these hot bearing detectors worked as designed, the system did not relay this information to the train’s engineers in time to prevent derailment. 

Early Detection via Acoustic Signals

Railways have long depended on temperature sensors to detect a bearing failure. As wear occurs, the heat due to friction will increase to the point where the metal becomes fatigued, and parts begin to break. Sensors are placed along the track to measure and detect unexpected temperature changes. Unfortunately, HBD couldn’t diagnose the problem until it was too late to prevent a disaster. Acoustic signals provide a faster and more accurate method for assessing bearing health. 

With this in mind, what if we listened to the train’s bearings instead of just measuring their temperature? Bad bearings don’t just get hot; they also create a distinctive and unpleasant sound. The sound of worn bearings is not always a sign of an imminent failure. Sometimes, a pump or other machine can operate with this sound signature for months before a system-stopping failure occurs. Using acoustic signals to detect this sound signature can aid in the early detection of mechanical failure.  

Application of Acoustic Technology

As we have seen from this incident, by the time a bearing failure is detected by heat signature, it may be too late to prevent an accident. However, when combined with machine learning, acoustic signals can detect signs of failure much earlier than heat detection methods. This early warning gives engineers the information they need to respond long before the point of failure. 

Another significant advantage of tracking acoustic signals over time is that this technology can follow the wear pattern of the part accurately over time. By training an algorithm on the acoustic properties of “normal,” it is possible to show deviations from normal that are not audible to the human ear. When machine learning classifiers are trained to detect subtle variations from normal, a relatively low-cost device can report accurate, timely results. 

There would be several ways to tackle this from an engineering and acoustics perspective. One method would be to place transducers along the tracks to listen to the bearings through the train’s wheels as it passes. Our research shows that the signal from worn bearings travels great distances with high fidelity through steel. 

While this approach is effective, the best application of acoustic technology would be sensors on the axle boxes of every car. Smart acoustic sensors built into the axle box of every train car could add significant safety margins for several different failure modes. In addition, acoustic sensors could give much more information than simply the bearing condition alone. 

In the case of this train derailment, the sensor on the failing set of wheels would have transmitted the condition of the bearings to the engineers, dispatcher, and the positive train control system, possibly days before the accident. According to video evidence, sparks and fire could be seen as the axle failed. This type of dramatic failure has an unmistakable sound signature that would be straightforward to detect using acoustic signals capture and analysis methods. 

Broader Applications of Acoustic Technology

In our experiments, we have seen that a single channel of an audible spectrum of audio captured from a mechanical system can reveal critical insights into the operation of that system. For example, a single acoustic sensor in the axle box could monitor the state of gears, bearings, brakes, and the track itself. Moreover, audio-centric machine learning tools can be trained to instantly detect deviations from ideal operating conditions to alert engineers of anomalies in various mechanical systems. 

The New Audio Machine Intelligence (AMI) Platform

BMT has developed the Audio Machine Intelligence (AMI) platform to address the need for integrated smart audio sensors. This system allows the bridging of acoustic and industry experts to create valuable solutions for a wide range of problems. So why aren’t we seeing more audio sensors on the market?  

Unlike visual AI systems, audio is esoteric and has not been a primary research focus in this sector. BMT seeks to eliminate any doubt that audio signals can be used in an industrial context by simplifying the hard acoustic problems with patented Digital Power Station (DPS) signal processing technology, novel user interfaces, and specialized audio capture hardware. AMI shortens innovation cycles to put the power of modern machine learning into the hands of engineers who need it the most. Instead of relying on pre-trained AI models, AMI provides an intuitive workflow for capturing data, training custom models, and deploying solutions unique to the user’s application.

Monitoring Machine Health with Acoustic Smart Sensors

Acoustic smart sensors offer an effective method for analyzing, capturing, and reporting machine health data. As we can see in the tragic example of the East Palestine derailment, temperature sensors and traditional inspections are insufficient for public safety. Temperature sensors cannot detect critical maintenance issues in time to prevent mechanical failure. We can prevent further catastrophes across many industries by providing high-performance reporting of fault conditions using audio analysis. The time is right to start bringing these smart audio sensors into production to improve and optimize many performance categories.