Tensor-Based State Scoring for Abnormal Process Detection

Instead of training the model to directly detect abnormal states using abnormal data, the patent inverts the approach by training the model to recognize normal states using only normal data. The abnormal state detection is then achieved by identifying deviations from the learned normal state patterns, effectively detecting anomalies through the absence of normal characteristics rather than presence of abnormal characteristics.

The patent extracts and focuses solely on the normal state data characteristics, separating them from abnormal state data. By extracting only the normal state patterns and building the model around these extracted features, the system eliminates the need to collect and process difficult-to-obtain abnormal state data while maintaining detection capability through deviation analysis.

The patent implements a feedback mechanism where the trained model continuously compares incoming data patterns against the learned normal state representations. When deviations are detected, the system can flag potential abnormal states, and this feedback loop allows the system to maintain high reliability by constantly measuring against the established normal baseline rather than requiring direct examples of abnormal states.

The patent performs preliminary action by thoroughly training the model on normal state data before deployment, establishing a robust baseline of normal behavior patterns. This preliminary training with abundant normal data prepares the system in advance to reliably detect abnormal states through deviation analysis, compensating for the lack of abnormal training data through advance preparation of the normal state reference.