Mechanical system anomalies may result in equipment failure that has major effects on
the environment, economy, and safety. Anomaly detection for mechanical equipment often demands
a great deal of domain expertise since many mechanical devices operate in harsh settings that make
them susceptible to many forms of failures. However, a typical problem in many real-world
applications is that one may not be able to get empirical information about anomalies or that the
historical data is entirely unlabeled, rendering traditional fault detection techniques inapplicable. This
paper fills the gap by proposing a novel deep learning-based method for anomaly detection in
mechanical equipment. It does this by combining two types of deep learning architectures, long short
term memory (LSTM) neural networks and stacked autoencoders (SAE), to identify anomaly
conditions in an entirely unsupervised manner. When the historical data is unlabeled and there is no
empirical information regarding anomalies, the suggested technique focuses on anomaly
identification using numerous feature sequences. The effectiveness and stability of the suggested
method are shown in an experiment for anomaly identification in rotary equipment using wavelet
packet decomposition (WPD) and data-driven models. The two phases of the process are LSTM
based anomaly detection and SAE-based multiple feature sequence representation. Fivefold cross
validation was used throughout the experiment to verify the stability and effectiveness of the
suggested method. The findings demonstrate that the suggested strategy may identify anomalous
working conditions with 99% accuracy in an entirely unsupervised learning environment and provide
a different way to use and incorporate characteristics for anomaly identification without the need for
empirical information.