Case Studies

Our client is a beverage bottling facility, processing hundreds of tin and glass bottles daily. The bottling facility itself is a conveyor belt with various stations each responsible for a specific part of a bottling process - arranging the bottles properly, pouring the beverage in, canning or corking the bottles, sticking labels on, etc.

Beverage bottling is a long process consisting of a series of sequential actions with each following action depending on the quality of the last one. If something goes wrong at any step of the process, not only the faulty bottle needs to be discarded, but it can affect the entire batch. For example, if a bottle for not properly sealed at a canning station, the overflowing beverage will interfere with the labeling process, resulting in either unreadable or unlabeled bottles.

Bottling line sampling, or hourly inspections of randomly picked bottles, is not effective nor efficient. It takes a significant amount of time to pick the bottles among dozens of containers, assess fill height, label placement, cork height, etc. Faulty bottles can easily slip by the inspector's eye, decreasing user satisfaction and resulting in a monetary loss for the company.

We were approached to develop a computer vision system that would use security cameras to detect faulty bottles at various stations, signaling when something went wrong. The system needed to work in real-time, analyzing security camera footage and looking for predetermined events. In addition, the system had to have a user-friendly interface comfortable to use for the average worker as well as provide real-time, daily, and monthly reports.

Utilizing security cameras that were already installed within the facility was not an option as high-quality video recognition required specific camera angles and lighting. The facility was equipped with a set of three security cameras which were installed at several stations along the conveyor belt and placed in a way that provided a clear view of the bottles. A dust and waterproof server was also installed to process the images captured by the cameras.

As our client has never used a monitoring system before and did not have video footage of the conveyor belt closeup, we were also tasked with the preparation and markup of a dataset for ML model training. The cameras were first used to collect footage and still images of both faulty and regular bottles, while our team analyzed the images and marked them up.

The ML model trained on the dataset is capable of detecting various events at different stations in real-time. The system monitors three conveyor belt stations - filler, canning, and labeling.

Filling lines - how much liquid is in a bottle - can be either too low or too high, depending on the nature of a mishap: a deformed bottle, degraded water seal, etc. The system can detect when a bottle is not filled properly and alert the personnel.

The canning process consists of two parts: a cap is placed on top of a bottle, after which it is pressed on to seal the bottle. Sometimes, due to liquid overflow, a deformed bottle, or equipment malfunction, the cap is not placed in a correct way or not placed at all, resulting in an unsealed bottle which affects the bottling process down the line - wetting down the labels making them soggy and therefore unreadable, or affect the beverage shelf life, resulting is a health risk for consumers.

Our system detects if the cap has been placed correctly and if the bottle is sealed properly.

The system can not only be used to detect faulty bottles but to assess the condition of equipment, detecting early signs of a breakdown. For example, bottles are usually filled low due to a degraded or broken water seal, which can lead to a complete production shutdown, causing significant monetary losses. Our system can detect these issues early on and let the client know the equipment needs to be inspected.

The system can also detect early signs of malfunction based on abnormal equipment moving patterns, increased vibrations, dirt or rust.

Daily and monthly reports help to diagnose whether the number of faulty bottles increases over time and which stations produce the most rejects. The client can easily access these metrics through a user-friendly dashboard both on mobile and desktop.

We have received overwhelmingly positive feedback from our client after system implementation. The monitoring and predictive maintenance system is fully operational and has already proven to be superior to standard quality assessment procedures. Predictive maintenance has helped our client to detect early signs of equipment wear down and react in time before a malfunction happens. Equipment idle time due to breakdowns has been lowered from 3% to 0,1%, and the amount of rejected batches has been decreased to 0.