The reliability and safety of high-traffic urban infrastructure elevator systems heavily rely on predictive maintenance. This paper proposes a method to identify mechanical failures in elevator door systems based on unsupervised learning methods. Based on a publicly available dataset of the elevator predictive maintenance project by Huawei, we model the time-series data of vibration, humidity, and door bearing sensors sampled at a rate of 4 Hz. The proposed pipeline combines classical models of outlier detection, such as Isolation Forest, Local Outlier Factor, and One Class SVM, with a deep learning-based LSTM autoencoder. The anomalies are identified through calculating reconstruction errors and comparing these values with a dynamic threshold that is formed based on training error distribution. The anomalies identified correspond with high-vibration oscillations that can be signs of mechanical anomalies, e.g. door jitter or control instability. The model obtains the capability to isolate abnormal events without fault labels at the start through visualization of both full and zoomed sensor timelines. The hybrid model is a powerful and understandable model which is easily deployable in the real-time in the elevator repair systems, integrating the advantages of both classical and deep learning frameworks. The paper shows the importance of AI in shifting towards proactive instead of reactive maintenance, minimizing down time, and increasing equipment life in electromechanical systems. The methodology can be widely used on other mechanical platforms that can monitor sensors.