Uporaba globokega učenja za izboljšanje kakovosti farmacevtskih izdelkov

POTOČNIK, BLAŽ (author), Begeš, Gaber (mentor)

Kakovost nas spremlja v vseh industrijah in povsod je preverjanje kakovosti izdelka ena najpomembnejših nalog proizvajalca. Razlogi za to so številni. Na prvem mestu je varnost uporabnikov. Podjetje lahko pride na slab glas, če bi nekakovostni kosi prišli do končnih uporabnikov in podobno. Praktično v nobeni drugi industriji ni odgovornost za varnost uporabnikov večja kot v farmacevtski. V preteklosti je farmacevtska industrija želela pregledovati večinoma tablete in trde želatinaste kapsule. V zadnjih letih pa so na priljubljenosti pričele pridobivati mehke želatinaste kapsule, ki zaradi razlik v pojavnosti in novih primerov napak, ki se pojavljajo na njih, otežujejo uporabo metod, ki smo jih uporabljali za tablete in kapsule. Ob naraščajočih zahtevah strank ne moremo več doseči njihovih pričakovanj po kakovosti pregledovanja. Z razvojem globokih nevronskih mrež v zadnjih letih, se je pojavila nova rešitev, ki jo zaradi principa delovanja lahko uporabimo za pregledovanje mehko želatinastih kapsul. Cilj mojega dela je bil definirati težave, ki jih mehko želatinaste kapsule predstavljajo pri pregledovanju kakovosti in ugotoviti, če so globoke nevronske mreže ustrezna rešitev za težave, ki so se pojavile z novimi oblikami farmacevtskih izdelkov. Ugotoviti želimo, ali lahko z njihovo uporabo izboljšamo kakovost pregledovanja, torej zaznamo več izdelkov z napakami. To posledično pomeni tudi izboljšanje kakovosti izdelkov, ki pridejo do končnega uporabnika. Da bi ugotovili, ali so globoke nevronske mreže rešitev, smo skupaj s strokovnjaki za razvoj algoritmov določili postopek priprave podatkovnih baz za učenje. Postopek testiranja je bil enak kot za rešitve, ki ne uporabljajo modelov globokega učenja. Za prikaz rezultatov sem izbral mehke želatinaste kapsule posebne oblike, ki jih brez uporabe globokega učenja ne bi mogli pregledovati s tako visoko kakovostjo. Izvedel sem primerjavo med rešitvama brez uporabe globokega učenja in z uporabo globokega učenja. Na teh izdelkih sem izračunal kakšna je razlika v kakovosti izdelkov po opravljenem pregledovanju. Za konec sem prikazal še uporabo globokega učenja na trdih kapsulah. Na treh primerih sem prikazal, kako z uporabo globokega učenja lahko rešujemo težave, ki se pojavijo zaradi razlik v izgledu mehko želatinastih kapsul. Z izračuni sem prikazal tudi, za koliko se izboljša kakovost zaradi uporabe globokega učenja.

• Language: Slovenian
• Keywords: globoko učenje, strojni vid, kakovost, mehko želatinaste kapsule
• Work type: Master's thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2022
• PID: 20.500.12556/RUL-142745
• COBISS.SI-ID: 130717699
• Publication date in RUL: 23.11.2022

Secondary language

• Language: English
• Title: Using deep learning to improve the quality of pharmaceutical products

Abstract

Quality is present in every industry and verification of quality is one of the most important tasks in production. There are numerous reasons for that but one of the if not the most important is making sure that end customer is safe. And in no other industry is safety of the customers as important as in pharmaceutical. In the past tablets and hard gelatine capsules were products that were important in pharmaceutical industry. But in recent years soft gelatine capsules have started to gain popularity. Soft gelatine capsules have tendency to have varying looks and have also introduced new types of defects. All that makes inspecting them extremely challenging using methods commonly used in inspection of tablets and hard gelatine capsules. With recent development in the field of deep neural networks, we are confident it is solution for the problem. The goal of this Thesis was to define problems that arose with inspection of softgels and find out if the use of deep neural networks could help us overcome those problems. Additionally, I also analysed by how much we can improve quality of products that end up in hands of end customer. To evaluate if deep neural networks present adequate solution, we, under leadership of experts for development of deep learning algorithms in Sensum d.o.o., established process of preparation of databases used for training purposes. Testing procedure was the same as for solutions used for tablets and hard gelatine capsules but with added focus on stability since technology was new. To present results, I used round soft gelatine capsules, where customer requested such high defect detection rate that it could not be achieved without deep learning algorithms. I made comparison of solutions with and without deep learning algorithms on same products and calculated difference in quality of products that reach end customer. At the end I presented how deep learning can also help us solve problems on hard gelatine capsules that arise because manufacturers search for improved procedures in production. On three cases I managed to illustrate how Sensum d.o.o. uses deep learning algorithms to combat problems that arose with popularity of soft gelatine capsules. Results presented in this Thesis show much improved quality of products after inspection with the use of deep learning algorithms compared to methods that were in use before.

• Keywords: deep learning, machine vision, quality, soft gelatine capsule