Embedded device-based autoencoder algorithm for anomaly detection in mechatronic systems

With the increasing application of computer technology in the field of electromechanical system operation and maintenance, accurate identification of anomaly features in multi-source sensor data and efficient model deployment in embedded environments have become key challenges for intelligent detection. To address the issues of redundant multi-source signal features, the inadequate adaptability of traditional detection algorithms to embedded environments, and weak noise immunity, this study constructs a novel anomaly detection model by integrating an attention mechanism, a sandcat swarm optimization algorithm, and a lightweight multi-layer autoencoder. Experimental results show that in the relevant data set, the area under the subject's working characteristic curve of the research model is 0.94, and the F1 value in fault detection is up to 85.2%. In low-computing embedded scenarios, the proposed model uses only 152.8KB of memory and has an inference latency of 32.5ms. In high-noise environments, the proposed model achieves an F1 score of 78.6%, significantly outperforming current mainstream comparison models. This proposed model effectively improves the accuracy, embedded adaptability, and noise immunity of electromechanical system anomaly detection, providing a new solution for enabling intelligent operation and maintenance of electromechanical systems with computer technology.