Fault Diagnosis of Semiconductor Detectors Using a Radial Basis Function Neural Network with Bayesian Optimization and K-Means Clustering

This study evaluates radial basis function neural networks (RBFNNs) for severity classification within predefined semiconductor-detector fault categories. The datasets contain simulated PN aging, RC circuit faults and radiation damage, together with three hybrid-fault constructions. Raw waveforms were partitioned before fitting feature transformations. A four-variant ablation over ten stratified partitions assessed Bayesian optimization (BO) of the hidden-layer size and radial basis width, and K-means placement of the centers. With training-fitted principal component analysis (PCA), the combined model achieved mean test accuracies of 97.67 ± 1.97% for PN aging and 100.00 ± 0.00% for RC faults, but only 7.78 ± 1.40% for the 66-class radiation task. An exploratory radiation reassessment using pulse-amplitude and residual-noise features, with a distance-scaled RBF width, achieved 72.42 ± 1.57% accuracy and 72.26 ± 1.63% macro-F1 over three partitions. Three frozen radiation models achieved a mean accuracy of 71.83% on 66,000 newly generated waveforms. Three PN aging–RC partitions gave 86.40 ± 1.52% accuracy within the standard search range. A separately selected expanded-range model achieved 96.33% on 110,000 new simulated waveforms. And three PN aging–radiation damage and RC–radiation damage partitions gave 77.40 ± 1.38% and 78.45 ± 1.28% accuracy with the standard search range, respectively.