Acoustic monitoring and control of Wire Arc Additive Manufacturing using Hidden Markov Models and reinforcement learning

Wire Arc Additive Manufacturing (WAAM) is prone to instability, spatter, porosity, and geometric irregularities that degrade the quality of the part. This article presents an acoustic monitoring and adaptive control framework that combines probabilistic process state detection with reinforcement learning (RL)-based corrective actions. Two broadband transducers recorded acoustic signals during the deposition of a functionally graded wall comprising 42 LA100S layers followed by 36 ER2209 duplex stainless steel layers (78 total). After cleaning and normalization, the frame-level spectral and cepstral features (RMS, spectral centroid, bandwidth, rolloff, MFCCs, and band-energy ratios) were extracted and reduced by principal component analysis (PCA). A Gaussian Hidden Markov Model (HMM) trained on these features identified recurring process regimes and their transitions, with states grouped into four conditions, stable-arc, unstable-arc, spatter, and porosity-like, to produce a layer-wise stability score. A two-component Gaussian Mixture Model (GMM) sets the stability threshold automatically, without manual tuning. An ɛ ɛ -greedy bandit controller then learned condition-specific corrective actions and issued layer-level recommendations. The results show clear regime transitions and a conservative RL policy that favors no intervention for stable layers and speed reduction for porosity-dominated layers.