One-Class Anomaly Detection for Zero-Label Fuel Injector Fault Screening at New Product Introduction

Engineers face significant challenges when a new fuel injector variant enters production. An automated bench test system can only evaluate what it is explicitly programmed to check. Conventional test benches, such as Hartridge systems, verify injector parameters against static OEM tolerances. However, this approach frequently fails to detect subtle deviations that are signatures of early field failure, contamination-induced back-leakage, and misfire events that dominate warranty returns. This challenge is most acute during new product introduction as no historical fault data exists to train supervised classifiers. To best our knowledge, this paper presents the first application of one-class anomaly detection to Common Rail fuel injector test bench data, validated on 1,000 production units from an automotive Injector manufacturer, where all units passed conventional threshold screening, yet 2.6% subsequently failed in service. This paper is structured as follows: Section I provides an introduction, Section II reviews the relevant literature, Section III describes the setup of the experiment and dataset, Section IV presents the four one-class model architectures, Section V reports results, Section VI discusses the hybrid deployment framework and economic analysis, and Section VII concludes.