Abstract: The adaptive extraction ability of process neural networks is used for dynamic time-varying signal's model characteristics. Based on this, a semi-supervised competitive learning process neural network with self-adaptive network structure is proposed to recognize patterns in an indicator diagram. Generalized discrete Fréchet distance is used as a metric between time-varying samples, discrete time sequence of the load and displacement is used as direct inputs, and the reward-punishment rule for updating is used to realize effective clustering under the constraint of labeled samples. The dynamic reconstruction of competitive nodes is realized according to the learning objective function of the network to reduce the influence of the initial cluster number. Experimental results show the effectiveness of the model and algorithm.