Case Study – Automotive Manufacturer

One of the major Asian automotive manufacturers wanted to explore Artificial Intelligence (AI) transformation for their factory floor to see if Artificial Intelligence could improve engine defect detection, and also track hard to read frame numbers as an automobile progressed along the assembly line.

This case study describes how the mSense product Sixth Sense helped them to improve their quality and streamline their manufacturing flow and in the process, there are no more engine assembly inspection headaches.

1. Situation- Engine Manufacturing

One of the major Asian automotive manufacturers, which makes around one million engines per year across four plants with 11 assembly lines/plant, wanted to improve:

  •  Quality Assurance (QA) by using  the running engine’s sound to more thoroughly detect manufacturing  problems and
  • Engine tracking through the manufacturing process, by automatically reading the frame number.

The current process was manual, labor intensive and error prone, and was not sufficient to keep up with the forecasted run rate. Customer needed a solution which improves their safety, manufacturing productivity and engine quality assurance.

2. Environment Challenges (Factory floor)

Customer’s manufacturing assembly/factory floor is compact and very noisy. There are multiple assembly lines and test stations in close proximity (10-12 feet) with engines constantly running. Each of the test stations has very limited space, and any changes to the assembly line are difficult and time consuming.  

The final test station, starts the newly manufactured automobile, sets the throttle, and a technician listens to the running engine. An experienced technician can usually hear a bad engine. 

To not impede the manufacturing run rate, a presumed bad automobile is pulled off the assembly line for further QA lab testing.   The technician needs to develop a keen ear to detect potential defects on the assembly line where several different types/models of engines with different fuel systems, exhausts etc.,  are assembled and tested. This experience is tough to gain and was not scalable.

3. Quality assurance and the manufacturing process

Unfortunately the technician listening to the engine’s sound could not catch all the bad engines.  In a total of about 2000 engines manufactured every day, post-manufacturing statistical analysis revealed that around 50 bad engines were slipping by the final inspection test.

Furthermore, to track each engine the manufacturer recorded the QR code. However, to eliminate confusion, the metal stamped 17 digit engine id number was preferred.Unfortunately, this id number was often in a poor location to read making it difficult for the technician.

4. mSense Sixth Sense Solution

Working with the manufacturer, mSense used its Sixth Sense product to train two purpose-built Deep Learning models to deploy on their factory floor:

  1. Acoustic classification inference model: To classify engine sounds with background noise and
  2. Character recognition inference model: To read an automobile’s Frame number in low light and varying backgrounds

Prototypes were developed within 3 months.

4a. A Sound Solution: 5X QA improvement

The Sixth Sense trained acoustic inference model improved QA by 5x detecting the bad engines that currently were missed.  A small footprint “non-invasive” hardware of 4 microphones for each final test station was able to quickly within seconds automatically detect the approximate 50 bad engines/day.

This level of accuracy was accomplished in spite of the background factory noise.   Now the technician while adjusting the throttle sees a green or red light for a good or defective engine. Furthermore in real time a QA dashboard as shown in figure 1 can be used to monitor manufacturing and spot trends before they become expensive problems.

Figure 1. mSense factory floor automatic acoustic inspection showing 62 good and 5 bad engine

Furthermore, mSense Sixth Sense inference models accommodate each of the type/style of engines simplifying QA across all engines.

Please see mSense demo video for more details.

4b. Frame number recognition: 6X performance

The customer for tracking purposes wanted to read the stamped frame number rather than just the barcode.  Numerous previous efforts to automatically capture the metal stamped engine id as shown in Figure 2, proved to be difficult because of the poor lighting, location and number of digits. 

Figure 2.  A portion of the metal stamped ID number image

mSense used a hand-held camera (similar to QR code reader) to capture the frame number’s image. Training a deep learning inference model, the image was translated to digits at a 99.0% confidence level: much better than the current manual technique. Furthermore to capture the frame number before took over 30 seconds while mSense’s solution took less than 5 seconds.

For the first time, the manufacturer had a safe and productive way of capturing this critical id number.

6. Summary: no more headaches

To summarize, mSense’s acoustic solution was able to:

  • Remove the loud factory floor background noise to “listen” to just the running test engine
  • The acoustic pass/fail test took seconds to complete
  • The test is automatic: no technician training is needed
  • The required hardware is a very small footprint/low power and is non-invasive requiring little changes to the assembly line

Our customer continues to engage with mSense in deploying these solutions across their assembly lines to accelerate their digital AI transformation resulting in improved factor floor productivity and QA.

 Over time with more automatically captured data and inference model refinement continued improvement in  engine quality will realize even higher product margins with no further headaches

For further details please contact us @ hello@msense.ai