Packaging evaluation is an important process to ensure the safety ofpackages. The process involves evaluating the packaging of the product to make sure it reaches the customer in a good condition. Packaging engineers evaluate the packaging design by using different methods such as mechanical testing, computer simulations like FEM, and collecting data from participants to simulate the real environment. Despite the value of data collection and mechanical tests in initial design, there remains a need for a fast and inexpensive method to continuously evaluate packaging designs post-implementation. While mechanical testing and computer simulations have been implemented to fulfill this need, these methods have limitations in effectively evaluating packaging performance. By increasing the abundance of data collections, data-driven approaches such as artificial intelligence (AI) have gained more attention. In this work, an attempt was made to implement AI in packaging evaluation. The implementation was examined in two phases: 1. A Novel Packaging Evaluation Method using Sentiment Analysis of Customer Reviews. As mentioned, current approaches toward evaluating packaging design (test and FEM) have their own limitations. These methods try to simulate the actual environment, but since these simulations cannot yet perfectly replicate real-life scenarios, the question remains of whether they are successful in this pursuit or not. Therefore, it is essential for the packaging engineers to evaluate their packaging design even after it is implemented. To address these shortcomings, a method was proposed to evaluate the packaging performance in an actual environment through customers’ reviews using natural language processing (NLP). NLP is a branch of AI that teaches computers to understand and process human language, enabling tasks like translation and sentiment analysis. Based on the results, engineers can identify potential sources of design failure. Moreover, the percentage of failures over various months and years is examined to identify the potential effects of seasonal changes on packaging failure. In our work, we compared three different TVs, labeled as A, B, and C. The percentage of packaging failure for each of them was 5.73%, 9.60%, and 11.19%, respectively, which means TV C had the worst protection through distribution. By using this method, the packaging performance can be evaluated from customer reviews instead of physical testing which will save time and cost. 2. Machine Learning Modeling of Patients’ Attention of Over-the-counter Medication Label Design. Progress has been made toward creating a model that can be used to study the effects of different parameters on label design and predict the patient’s behavior with respect to parameter changes. In this respect, data provided by the Healthcare, Universal Design, Biomechanics (HUB) research group at Michigan State University (MSU) was utilized. They implemented a new design for Over-the-Counter (OTC) medication. They put a small box in front of the package (FOP), highlighting important information for the patients to increase patients’ attention, and their likelihood to read the label of OTC medication. The HUB research group gathered their results from ninety-two participants by using a change detection method. The goal of this project was to predict the response time of the participants with respect to changes. In this project, the impact of content change on response trials with three classes (hit, miss, and time out) was studied using three ML approaches: Decision Tree (DT), Random Forest (RF), and K-Nearest Neighbors (KNN), chosen for their effectiveness with categorical data. The models were trained and tested for accuracy and area under the curve (AUC), which measures their ability to differentiate between response classes. Results showed accuracies of 85% for DT, 88% for RF, and 90% for KNN, with corresponding AUC values of 0.5, 0.54, and 0.5. Despite satisfactory accuracies, the AUC scores indicated that the models’ performances were not as effective as expected. Further analysis indicated that the models’ limitation to predicting only in the ’hit’ class was due to imbalanced data. By applying SMOTE, the dataset was balanced more effectively, resulting in a boost in model performance. Although the accuracy decreases to 78% and 81% for RF and DT models, respectively, the AUC value increased to 0.81 and 0.79 for RF and DT, demonstrating SMOTE’s efficacy in managing imbalanced datasets. The model significantly improved its classification of the ’hit,’ ’miss,’ and ’time out’ classes.
Read
