Unsupervised neural network for fault detection and classification in dynamic systems

We present recent results of using a Kohonen neural network to detect and classify faults occurring in a dynamic system. The measured outputs from the system are first used in a Kalman filter to generate residual vectors that serve as fault indicators. As the residuals are generated they can be sent one-by-one to the Kohonen network, both the Kalman filter and the Kohonen network operating in real time. The Kohonen network detects and categorizes the fault, since the residual vectors serve as signatures for different types of faults. The Kohonen network starts with a few pre-designated categories, each category representing a fault type. As more and more residual vectors become available, the Kohonen network opens new categories for residuals that do not have a good enough match with any of the existing categories. The concept is illustrated by an application example that uses actual fault data commercially recorded by the utilities in Texas.