Lifetime testing in the durable home appliances sector is a time and resource-intensive validation process, especially for tumble dryers that require 1600 operating cycles to represent the estimated product lifetime. In practice, intermediate decisions are often made around the 320th cycle to accelerate product development; however, these decisions largely depend on expert judgment and may fail to detect failures that occur in later cycles. This study proposes a data-driven Predictive Quality (PdQ) framework to support early and more reliable evaluation of tumble dryer lifetime tests. Due to the limited availability of faulty samples and the difficulty of identifying exact fault onset points in long sensor sequences, an unsupervised deep learning (DL) approach was adopted. An LSTM-based autoencoder (AE) was trained using healthy C-class machine data to learn normal operating behavior from multivariate time-series signals, including power consumption, temperature variables, and control signals. The trained model was then evaluated on E-class machine data containing both healthy and faulty test cases. Reconstruction error was used as the anomaly score, and a multi-level thresholding structure was applied to classify test behavior into normal, warning, and critical risk levels. The model achieved an F1-score of 82.35%. The proposed workflow includes data cleaning, preprocessing, model training, anomaly score generation and decision support interpretation. The results indicate that the proposed approach has the potential to generate early risk indicators, transform raw test data into actionable insights, support operator decisions, and improve the efficiency of lifetime testing processes.

Keywords

Predictive Quality, Deep Learning, Decision Support System, Anomaly Detection, Lifetime Testing.