Employing Artificial Neural Networks to Judge Vehicle Vibration Quality

By Jennifer Bastiaan, Kettering University and Ed Green, Hottinger Bruel and Kjaer

Ed Green of Hottinger Brüel & Kjær (HBK) and Jennifer Bastiaan of Kettering University, have been collaborating on an automotive vibration quality and ride comfort project. The ride comfort of ground vehicles is associated with vibration quality as perceived by the human occupants of the vehicle. However, it is challenging to relate human perception of vibration quality to measured vehicle responses. Ed and Jennifer are in the second phase of a two-phase project on the topic.

Phase 1: Human Intelligence 

In the first phase, physical measurements from inside four vehicles traveling over a cleat were used as inputs to a COMPACT Simulator from VI-grade. The test vehicles included two passenger cars, a 2012 Ford Focus and a 2017 Chevrolet Bolt, one SUV, a 2020 Jeep Wrangler, and one pickup truck, a 2021 Dodge Ram 1500. Cleat testing was performed at the GM Mobility Research Center (MRC) proving ground, located on the campus of Kettering University in Flint, Michigan. 

The four test vehicles traveled over a rubber cleat on the driver’s side. Three vehicle speeds (10, 15, and 20 MPH) were tested along with three tire inflation pressures (15, 32, and 44 PSI). As there were nine cleat tests for each vehicle, and four vehicles were tested, there were 36 cleat test events that were measured for use with the COMPACT Simulator. Instrumentation in the vehicles included: 

• Steering wheel accelerometers at 3 o’clock and 9 o’clock (lateral and vertical directions) 

• Heel point accelerometer embedded in a custom floor mat (lateral and vertical directions) 

• Seat accelerometers on the driver’s seat cushion (lateral and vertical directions) and the seat back (longitudinal direction) 

• Binaural microphone headset in the driver’s ears (left ear and right ear sound pressures) 

• Action camera on the driver’s sun visor (forward-facing view) 

Only the seat cushion vertical acceleration, the driver’s ears sound pressure measurements, and the action camera video were used with the COMPACT Simulator, to simplify the setup. A human jury of six people, half women and half men, were tasked with rank ordering the cleat tests using a paired comparison procedure. The 36 cleat test events were ranked from 1 (best) to 36 (worst). In general, the jury preferred the low tire inflation pressures, ranking the cleat test events with low tire pressures as less harsh than the events with high pressures. This was an expected result.  

The unexpected result was that the jury generally preferred the utility vehicles over the passenger cars. This was unexpected because SUVs and pickup trucks are designed for function over comfort. Generally, lower seat cushion peak vertical vibration levels were associated with better rankings. The results of the phase one project were published in the paper “Vibration Quality and Ride Comfort Investigation with Transient Excitation in a Ground Vehicle Simulator Environment” and presented at the May 2023 INCE-USA NOISE-CON event in Grand Rapids, Michigan. 

Phase 2: Artificial Intelligence 

The goal for the second phase, which is still ongoing, is to relate the opinions of the human jury to the physical measurements using artificial intelligence. Typically, the vibration quality or ride comfort of a vehicle is evaluated by a human subject matter expert, who offers opinions based on a Likert-type scale. These scales often include qualitative descriptors such as “ridiculous” and “perfect”. While human experts can provide valuable insights, we want to know if physical measurements can be automatically processed using artificial intelligence to produce opinions similar to the human experts. 

To create the mapping task for the artificial neural networks, the jury rankings from 1 to 36 were divided into five vibration quality categories:  

1) Excellent (Rank 1 to 7) 

2) Good (Rank 8 to 14) 

3) Fair (Rank 15 to 22) 

4) Poor (Rank 23 to 29) 

5) Very Poor (Rank 30 to 36) 

These categories and the associated physical measurements were used to train and test the artificial neural networks. Specifically, Multi-Layer Perceptron (MLP) networks with backpropagation learning were employed to solve the classification problem using MATLAB. Two separate data sets were collected on the MRC proving ground in November 2022 and October 2024. The second data set was necessary as the first data set did not contain enough training data for the networks.  

Seat cushion vertical acceleration data, downsampled to 8192 Hz, was used for training the networks. Modifications were made to the number of neurons and hidden layers in the networks. Different training algorithms and transfer functions were investigated for the networks. Classification was mostly successful, as multiple networks studied could identify four out of five categories. Interestingly, many of the MLP networks had trouble mapping measurements to the second category of “good” data, confusing it with the third category of “fair” data. The networks had something in common with the human jury, therefore, as the humans had some trouble discerning between middle-performing vehicles, sometimes asking for the same vibration pair to be played over again many times. 

The MLP networks did a reasonably good job predicting human opinions based on vibration measurements. Future work on this project will include experimenting with other artificial neural network types and adding sound pressure measurements into the training data, as the human jury was presented with the measured soundscape and it is only fair if the networks are presented with the sound pressure data as well. Furthermore, the networks will be trained on subsets of the measurements, such as peaks, valleys, and overall levels. Finally, the statistical variation of measurements in multiple test runs will be investigated, as many runs of the “same” test produced curiously similar, yet different, results. 

The outcomes of the phase two project will be presented orally at the May 2025 SAE NVC event in Grand Rapids, Michigan, with the title “Mapping of Vehicle Vibration Quality Measurements to Human Expert Opinions Using Artificial Intelligence”. Ed and I hope to see you there at the event. We love to discuss vibration quality, as it remains a fascinating subject with unanswered research questions.