Laser Machining Signal Reconstruction for Unknown Anomaly Detection

The patent applies preliminary action by training the autoencoder exclusively on normal process signals before deployment. The autoencoder learns to reconstruct normal patterns during a pre-training phase, and during actual monitoring, any deviation from these learned patterns is flagged as an anomaly. This eliminates the need to collect and train on rare error examples, as the system is already prepared to recognize normal behavior and identify deviations.

The patent uses copying by creating a reconstructed version of the input process signals through the autoencoder. The autoencoder generates copies of the normal signal patterns during training, and during operation, it produces reconstructed signals that are compared with actual signals. Anomalies are detected when the reconstructed copy deviates significantly from the original, providing a mechanism for anomaly detection without requiring examples of the actual anomaly.

The patent applies dynamics by using a neural network-based autoencoder that can dynamically adapt to changing process conditions. Unlike static reference curves and envelopes, the autoencoder learns complex nonlinear patterns from training data and can adjust its internal representations when retrained with new normal process data. This dynamic capability allows the system to maintain high adaptability while preserving ease of operation through automated anomaly detection without manual envelope tuning.

The patent employs parameter changes by transforming the approach from fixed threshold-based detection to a learned representation-based detection. The autoencoder learns optimal parameter transformations during training, mapping input signals to a compressed latent space and back. This parameter transformation enables the system to adapt to varying process conditions by learning robust feature representations that remain stable despite changes in operating parameters.