Anomaly Detection with MEMS Vibration Sensors and Machine Learning - Part 1/3 Introduction

Scope and TL;DR

This post explores how wireless vibration sensors and machine learning techniques are used for anomaly detection in industrial shaker screens. By monitoring vibration data and using deep-learning models, various types of mechanical faults can be detected at an early stage.

This is the first part in a three-part series.
Part 2: Anomaly Detection with Autoencoders
Part 3: Real-world Vibratory Screen Data

We cover the technical setup and data postprocessing using autoencoders (part 2), which are the key elements for reliable and fully automated fault detection.

Finally, based on real-world data collected during an 7-month long data collection example from a household waste processing plant in the Benelux, the reader will understand both the strengths and weaknesses of the system (part 3).

This read is targeted at both the vibration expert and the casual reader interested in gaining better insight in the practical applications of machine learning, beyond the hype that has surrounded it in recent years.

Introduction

Waste management plants rely on long serial processing lines. Failures in any pivotal stage may cause severe capacity loss due to the limited redundancy. This makes some level of monitoring targeted towards machine health and predictive maintainance crucial to avoid unplanned downtimes and the cost that inevitably comes with it.

See “How does the post-separation process work?” by AVR.nl for additional photos.

Figure 3: Overview of the Household Waste Separation Process (click to enlarge) [credits: AVR].

While digital PLCs are in the control of the pipeline and can already detect the most acute faults in realtime, machine manufacturers treat predictive monitoring mostly as an afterthought, leaving it to the maintanance staff to handle unexpected random failures. However, in unhealthy environments filled with dust, humidity, and noise, routine manual inspections are not practical.

Figure 4: Unhealthy environment at a waste processing plant —
Accumulation of dust on the receiver module after several weeks of operation.

In this blogpost, we will show that ruggedized wireless vibration sensors and machine learning in the data postprocessing chain provide us with consistent health updates at a regular interval, enabling early detection of issues. We shall cover the technical setup, the data processing, and the machine learning which automate the anomaly detection in noisy environments.

Noisy, in this case, refers to the sensor signal, which is polluted with unwanted vibrations from the production process itself, in addition to the machine defects we are trying to detect here.

---

Monitoring Vibration Screens

Monitoring vibrating conveyor equipment in waste management plants presents unique challenges. These plants are large, and network infrastructure or good cellular coverage is often lacking. Installing cabling across large sites is not only expensive but also prone to failures. This is especially true for machines that experience significant vibrations.

One example such an is a vibratory screen, which pre-sorts the household waste material on size before it undergoes optical sorting. In the optical sorter, spectral cameras detect different materials, and pressurized air is then used to further separate individual pieces into different output flows.

See “SPALECK: Recycling Waste Screens” [spaleck.eu] .

Figure 5: Vibratory Screen with the location of the Vibration Sensor [image credit: SPALECK].

These vibratory screens, driven by a synchronous motor that powers an eccentric axle with counterweight, pose several challenges:

• Large Displacements: Translational displacements of up to 10 cm make not only the shaker itself prone to considerable stresses, but also any wired sensor setup will be susceptible to both connector wear and cable fatigue in the long term.

• Regular Maintenance: Parts of these machine are frequently disassembled for maintenance. It is preferred that any sensors should be as non-intrusive as possible during such manipulations. This ensures a consistent sensor location and orientation and improves the accuracy of the machine learning algorithms.

Given the challenges of monitoring shaker machines, the customer here opted for wireless sensors to avoid the above issues.

figure 6: Video Demonstrating Sensor Placement and the Challenging Conditions on a Shaker Screen in a Waste Processing Plant.

To ensure that routine maintenance tasks can be performed without disturbing the sensors, they were installed 40 cm away from the ideal location.

The sensors are located on the machine frame, instead of directly on the bearing housing of the eccentric axle, which inevitably reduces the sensitivity to high-frequency bearing faults. This highlights the gap between ideal lab conditions and a typical installation.

So factors that need to be considered include:

• Signal Propagation: The sensor’s non-ideal placement introduces multiple wave propagation boundaries (i.e. transitions between materials), which attenuate and reflect high-frequency signals. This makes any measurements above 10 kHz, including ultrasonic frequencies, costly and of limited benefit.

• Unsuitability of RMS Sensors: For simple RMS sensors, the fault signals will be masked by process noise. This makes them ineffective for vibratory screens except for the most obvious late-stage catastrophic failures (see part 3).

---

MEMS Accelerometers

Given the above prerequisites, triaxial MEMS-based wireless vibration sensors are an ideal choice.

Figure 7: Microphotograph (2x2mm) of a MEMS Accelerometer (without the ASIC processor).
Hollocher et al., "A Very Low Cost, 3-axis, MEMS Accelerometer for Consumer Applications," 2009. [researchgate.net] .

Fully integrated MEMS vibration sensors (micro-electromechanical system) detect acceleration by measuring changes in capacitance (distance) between a fixed electrode and a suspended on-chip proof mass. The variations in capacitance are then digitized by the ASIC postprocessor embedded in the same SiP package and converted into their corresponding acceleration values.

Characteristics of MEMS:

• Robustness: Highly durable and well-suited for long-term use in harsh environments. Piezo accelerometers on the other hand, have superior noise characteristics but are more prone to material failures of the ceramic sensing elements and need more careful handling to maintain their calibrated sensitivity.

• Frequency Response: MEMS sensors capture vibrations up to a few kHz(*) in three axes. This allows for spectral separation of process noise and fault frequencies, so we can tune for the optimal the signal-to-noise ratio and fault sensitivity.
  (* the installed sensors have a Nyquist -3dB bandwidth of +/-1KHz)

Figure 8: Peak and Noise Performance of the iQunet IVIB161010-ACC3-016 Accelerometer.
Detailed background on Frequency Response and RBW performance figures here .

Advantages of wireless sensors:

• Wireless Range: With a range of 20 to 50 meters, wireless sensors eliminate the need for fragile and expensive cabling. Downside is the need for battery replacement every 20-50,000 spectral measurements.

• Cost: Each vibratory screen is equipped with 4 triaxial sensors. Switching to piezoelectric sensors would significantly increase costs, tripling the expense for the sensor elements alone compared to the total cost of the complete wireless MEMS-based system.

Figure 9: Despite offering better sensitivity, a wireless IEPE bridge, with its additional cabling, is not a viable alternative to the existing 4x wireless MEMS triax sensors on the vibratory screen.

In the next chapters, we will discuss how the sensor data is post-processed and used to align otherwise unplanned downtime with scheduled maintenance tasks, effectively reducing standstills to virtually 0 excess downtime.

---

Part 2: Anomaly Detection with Autoencoders