Source: Applus IDIADA

ADELANTO, Calif. – The following is the third of three Technical Corner pieces by Antonio Rubio, Project Manager, Braking Systems at Applus IDIADA, on how the use of artificial intelligence can help in the development of braking systems NVH characteristics.

The first article:

Brake NVH Development Process Improvement (Part 1 of 3)
Brake NVH Development Process Improvement (2 Of 3)

This article presents how artificial intelligence can support the identification of brake noise in real time.

Particularly, the validation of the algorithm is shown in order to cover not only brake squeal in standard condition, but also to detect different brake noises, under different testing conditions (different standards, city and mountain driving, low and high ambient temperature, different vehicle category).

This third and last part of the article presents an automatic process which is managing the complete process, from driving and rating, through detection, brake noise automated analysis and finally the upload of the testing report and relevant information in a connected secure environment.

  1. Noise Detection & Algorithm Validation
    1. Brake Squeal Initial Validation

The model was initially trained with data from different vehicles mainly from Mojacar testing vehicles. Data owned by Applus IDIADA from internal testing vehicle was used in order to guarantee the repeatability of the detection in terms of driving style, acquisition system and sensor installation.

Figure 5 – Brake Squeal Validation and Detection by ML Algorithm

After the training phase, the work followed up with validation of squeal data. Results from four different vehicles were used, and detection using “traditional” techniques (combining subjective driver evaluation and objective detection) with the newly developed algorithm are compared. It is considered that the standard technique set the reference value for the detection.

Table 1 – Vehicle Testing Validation

It is clearly shown how the algorithm can detect brake squeal noise with a level of confidence at least comparable with an expert and trained master driver. The training of the algorithm was performed using data from Mojacar testing (mainly country road driving), while the validation was successful for both Mojacar routes and Barcelona City Traffic tests as well.

  1. 2. Main Validation Process

The reliability of a ML model comes from its ability of detecting/predicting on unseen data, i.e., data not used during the training process.

Besides that, one important question should be “can the model detect noise anomalies in a wide range of scenarios that can be different from the ones used for training?”

Using data collected from completely different scenarios (weather conditions, types of vehicles, etc), the model is validated.

In order to perform this final validation stage, which is considered to be the final sign off of the complete algorithm development process, data from different localizations (Mojacar, Barcelona, Southern Germany, Los Angeles, Northern Italy) is used.

A minimum of two or three routes per localization is considered and all the data has been validated by an expert engineer.

Different type of vehicles (SUV, Sport, C and D segments) are considered, including routes with rain and snow condition, but also data with different types of noises (squeal with different freq., wire brush, creep groan etc.); it has to be underlined that the current detecions is mainly focused on squeal. Off-braking noises and incorrect recording (simulated accelerometer failures) are considered as well.

The following matrix of configurations has been considered:

A total number of 4724 noisy events are considered for the final validation.

This final validation has been performed under the following assumptions:

  1. Squeal which is detected and validated by the engineer is categorized as “correct”
  2. Squeal not detected is categorized as “not detected”
  3. Other noises detected as squeal are categorized as “incorrect” (scrape, creep groan etc.)
  4. AI Algorithm has been trained with squeal noises, so only squeal has to be detected.
  5. AI Algorithm has not been trained to detect wire-brush/gurgle
  6. Squeal detected inside the wire-brush/gurgle by the AI is categorized as “correct”

As a summary, these are the results obtained by this final validation:

  • Correct detection: 95.2 percent
  • Not detected events: 4.2 percent
  • Incorrect detection: 0.6 percent
  • In addition to these:
  • In all the locations, the correct detection is higher than 92.3 percent
  • In all the locations, the not detected events are less than 6.8 percent
  • In all the locations, the incorrect detection is lower than 1.6 percent

    2. Testing Application

The final aim of this tool is to be applied in real testing condition in order to provide a reliable automated brake squeal detection. The final objective is to have a better detection with respect to the traditional triggering based on simple Boolean conditions, since this algorithms has been trained with data which has been entirely validated by human hearing.

In order to apply this technique, the following process has to be followed:

  • Step 1: Data is collected by a DAU during a brake NVH durability route in a defined environment
  • Step 2: A Process Manager Software has been developed in order to process the following steps, without human intervention:
    • Step 2.1: Transfer the data from the DAU to the remote server
    • Step 2.2: Perform the automatic brake noise classification through machine learning
    • Step 2.3: Finalize analysis and generate the corresponding route reporting
    • Step 2.4: Manage the data ingestion into the cloud
    • Step 2.5: Automatically transfer the data to the corresponding reporting tool.

In this way, a complete automatic process has been generated, from data detection to final reporting.

3. Conclusion

Convolutional Neural Network has proven to be effective for brake noise identification and detection.

Very high level of confidence is shown for brake squeal detection (always higher than 95% for squeal detection with respect to master driver subjective detection). A final validation under different conditions has also been proved.

Lower confidence is shown for other high frequency noise detection (wire-brush, gurgle, multi-frequency squeal, squeak).

Detection of high frequency noises should be improved; further algorithm training has to be carried out (these are noises with generally lower occurrence, that’s why a higher number of testing vehicles will be necessary). Thus, brake squeal detection algorithm will have to be validated again once the implementation of this alternative noises is included in the detection protocol.

Correlation with subjective assessment by master drivers is under investigation; the correct acceptance level has to be defined and is the next big challenged to be technically solved.


With more than 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 has locations in California and Michigan, with further presence in 25 other countries, mainly in Europe and Asia.