A Method for Constructing an ADVI-Based Statistical Discrimination Model for Hammering Inspection of Highly Mixed Cast-Iron Parts

摘要

This paper describes a method for constructing a statistical discrimination model to inspect defective cast-iron parts such as graphite-spheroidized defective products by hammering test used in multi-product mass production. The hammering-sound signal includes multiple resonance-line spectra those frequencies vary by defect. For constructing the model, only non-defective hammering-sound data that can be collected from the production line is input, and a distribution function that fits frequency distribution of each resonance-line spectrum is estimated. Since the frequency distribution shows multimodality and asymmetry, estimation is performed by using automatic differentiation variational inference (ADVI) with a mixed normal distribution function. A confidence interval of the obtained distribution function is then regarded as non-defective-product existence section, and the discrimination model is automatically constructed by connecting the sections of all the resonance-line spectra in the audible range. Then, products outside the sections are discriminated as defective products. Results of experiment confirmed that it is possible to achieve hammer-test inspection with detection rate of 100% and prevent overlooking of defective products.