The traditional fault diagnosis method based on classifiers relies heavily on complete prior knowledge of the existing faults and often requires complex denoising calculations. However, in practical applications, it is difficult to obtain the complete fault state data, especially degradation state data, in advance. This reduces the classification accuracy of unknown new states and severely limits fault diagnosis. To overcome these difficulties, a novel analog circuit fault diagnosis and unknown state recognition method based on density peak clustering and voting probabilistic neural network (VPNN) is presented. In this method, the novel VPNN model is constructed based on prior knowledge of faults, in which a majority voting algorithm is applied instead of the complex process of denoising, to restrain the environmental noise. Further, a newly designed discriminant layer is added after the summation layer to identify the unknown state data, pursuant to which a derivative-based method is proposed to eliminate the transition data. Moreover, to update the preliminary VPNN, the K-nearest neighbor and density peak clustering procedures are applied to automatically determine the number of new pattern neuron classes, and a data reduction algorithm based on the Gaussian mixture model is proposed to determine the pattern neuron samples. Accordingly, numerous redundant samples are reduced. Thus, new pattern neurons can be straightforwardly added to the preliminary VPNN. Our experimental results clearly demonstrate the robustness of the proposed method, which can reduce false alarms, identify unknown states, and determine the newly added pattern neuron classes and pattern neuron samples to update the VPNN automatically.