Article by: Antonio Rubio, Project Engineer, Braking Systems in Applus IDIADA
This series presents a study conducted in the area of brakes, aiming to replicate the assessment performed by an expert driver regarding the annoyance level of brake noise.
Squeal brake noise is a critical aspect of vehicle performance and can impact the overall customer experience. Currently, subjective assessments of brake noise are made by human evaluators, whose training and skilling can be time-consuming, while the results are still based on subjective criteria. By using the AI model to predict subjective ratings based on objective measurements, this process can be automated and made more consistent.
The project involved using provided data containing noise samples along with their corresponding ratings to develop a model capable of accurately predicting subjective ratings. The data analysis, development of initial models, refinement of variables, and evaluation of the final model are discussed in detail. The results demonstrate the potential of machine learning techniques to replicate driver evaluations of brake noise.
The paper provides an overview of the context and motivation behind the project, emphasizing the importance of accurately assessing brake noise levels. The objective is to develop a model that can replicate the subjective evaluations made by expert drivers.
The study utilized a comprehensive dataset collected during several years of testing at Applus IDIADA. Subjective ratings come from a selected highly skilled driver, and corresponding objective measurements from recorded data through typical brake durability instrumentation. Exploratory data analysis (EDA) was performed to examine the correlation between various variables and to identify any patterns in the data. The EDA showed that there was some correlation between certain objective measurements and subjective ratings. In addition, the data set was cleaned in order to remove outliers. Based on these findings, the most relevant variables were selected to be used in the model.
Data set description
The dataset comes from several years of testing at Applus IDIADA during Brake Noise Durability programmes performed in Mojacar (Almeria). Subjective ratings come from a reference selected highly skilled driver. The reason to use one reference driver is to minimize the possibility of criteria differences between drivers that could hinder training and modelization.
The data set contains data from several years of testing, with various vehicle segments, vehicle types (BEV, ICE etc.) and noise characteristics (frequency, duration etc.).
Subjective ratings from the driver through a rate box are linked to each noise. Subjective ratings are between 1-10, with 1 meaning worst rating and 10 no noise/imperceptible. The following table 1 defines subjective criteria used for squeal noise event:
Table 1 Subjective rating criteria
Objective data were recorded during testing from IDIADA internal data acquisition system dbBrake using typical brake durability instrumentation to characterize squeal brake noises such as frequency [Hz], amplitude [dB], corner origin from accelerometers mounted in the calipers [g] and conditions such as brake disc temperature [ºC], brake pressure [bar] and ambient temperature [ºC] and humidity [%]. Note that instrumentation used was two microphones in cabin, close to the position of driver’s ear.
Exploratory Data Analysis (EDA)
Once the data has been preprocessed to obtain the necessary information, the resulting dataset is subjected to exploratory data analysis (EDA).
During the EDA process, the dataset is analysed statistically to gain insights and understand the individual variables, their interactions, correlations, and other relevant characteristics. The goal is to explore the dataset comprehensively and identify any patterns or trends that may exist.
The initial data set consists of data of the selected driver, which corresponds to 622 routes. There are 115 variables analysed in the data set which could be checked in figure 1.
It was obtained the Pearson correlation coefficient between all variables. The Pearson correlation coefficient, also known as Pearson’s r or simply correlation coefficient, is a statistical measure that quantifies the linear relationship between two variables. It takes values between -1 and 1, where:
- 1 indicates a perfect positive linear relationship (as one variable increases, the other increases proportionally).
- -1 indicates a perfect negative linear relationship (as one variable increases, the other decreases proportionally).
- 0 indicates no linear relationship between the variables.
The results of the Pearson correlation between the initial 115 variables could be checked in figure 10 in the ANNEX of current paper.
Figure 1 Initial 115 variables of data
As an overview, 28.3% are missing cells and 5.9% duplicate rows.
After analyzing the dataset as a whole, it is determined that a data cleaning process can be performed to discard variables that do not contribute value to the model. This step involves removing variables that are deemed unnecessary or do not significantly affect the target variable, which is the rating of brake noise annoyance.
The purpose of this data analysis and cleaning stage is to prepare the dataset for model training and ensure that it contains the most relevant and informative features. By removing irrelevant variables, the subsequent models can focus on the essential aspects of the data and improve the overall performance and accuracy of the predictions.
Finally, after variables review from Brakes department and model results, it is decided to include the following variables in the dataset to train the model (figure 2):
Figure 2 Final selected variables for the model
From previous analysis, the Pearson correlation coefficients between the target variable to be predicted “Rate”, the subjective rating from driver, and the rest of the variables can be extracted (figure 11 in ANNEX of current document). The main variables with more Pearson correlation with “Rate” are (table 2):
Table 2 Main Pearson correlation coefficient with the target variable “Rate”
According to the Pearson correlation with target variable “Rate”, main correlation is with “Max mic”. In addition, there is some correlation with “Duration” and “Mic freq”. These results are coherent as “Max mic” represents the maximum microphone amplitude in decibels, “Duration” the duration in seconds and “Mic freq” the frequency of the squeal noise event.
It should be noted that positive correlation means that target variable increases if variable increases or vice versa. Negative correlation means that target variable increases if variable decreases or vice versa.
In addition, it is observed that there is also a strong correlation with “LH mic amp 1” that represents the amplitude of the left-hand side microphone close to the driver’s left ear and “RH mic freq 1” that represents the frequency of the right-hand side microphone close to the driver’s right ear. Variables “Max mic” and “Mic freq” extract the data from both “LH” and “RH” microphones, so it is coherent to have correlation with these variables too.
About Applus IDIADA
With over 25 years’ experience and 2,450 engineers specializing in vehicle development, Applus IDIADA is a leading engineering company providing design, testing, engineering, and homologation services to the automotive industry worldwide.
Applus IDIADA is located in California and Michigan, with further presence in 25 other countries, mainly in Europe and Asia.