Additive manufacturing (AM) today enables the production of components that are difficult or even impossible to fabricate using traditional technologies. It also helps reduce scrap rates (generally referred to as the buy-to-fly ratio) and the environmental impact associated with each part’s production. However, due to the complex physical phenomena involved, AM parts are often more prone to defects, leading to scrap, financial losses for companies, and increased production costs per part. For this reason, every AM product, particularly those intended for structural or safety-critical applications in metals, must undergo certification through non-destructive testing (NDT) equipment. However, these procedures are typically performed on random samples and are time-consuming. To address this issue, Dr Giulio Mattera from the University of Naples “Federico II” (Italy) and his colleagues, and Professor Zengxi Pan from the University of Wollongong (Australia) have developed a methodology that employs Artificial Intelligence to streamline the certification process, with potential benefits in terms of both cost reduction and product quality improvement.
“Data exhibit complex structures and relationships, especially in welding-based technologies,” explains Dr Mattera. “Therefore, a more sophisticated analysis of the data, enabled by the combination of advanced data analytics tools and Artificial Intelligence, can outperform traditional methods based on simple statistical descriptors such as the mean and variance of the process.”
The newly developed approach processes welding current and voltage data collected at high frequencies, above 5,000 samples per second, to extract information from both the time and frequency domains. This allows for the capture of not only the statistical descriptors of these process variables but also the structural patterns underlying their repetitive nature.
As Dr Mattera emphasises, “A stable welding process, such as the Wire Arc Direct Energy Deposition process, is characterised by the repetition of similar waveforms in both welding current and voltage, which are directly associated with the melting and deposition of the filler wire in the component.” For the research team, by jointly analysing information from both domains, it becomes possible to better assess process stability and identify any anomalous conditions that may be linked to defects, thereby preventing quality degradation.
“Another key innovation of this work lies in the minimal prior knowledge required to replicate the methodology across different industrial systems,” explains Dr Mattera. “Unlike conventional AI models that demand extensive datasets containing both good and defective samples, our approach only needs data from non-defective conditions. This drastically cuts down the time and cost of data collection. The model then learns to recognise normal behaviour and automatically flags any anomalies, effectively detecting defects without ever having seen one.”
The researchers carried out a series of experiments in which they varied several process parameters to collect data from high-quality deposition conditions. They then trained an AI model - specifically, an algorithm known as Isolation Forest - to learn the map of normal behaviour within the data. When tested, the algorithm successfully identified anomalous conditions and assessed the component’s quality. In one case, it detected irregularities that indicated the need to clean the welding torch nozzle during production, preventing the formation of defects such as porosity in the metal part. Such defects, if left unchecked, could compromise the mechanical properties of the final product.
“By comparing the proposed methodology with traditional Statistical Process Monitoring (SPM) techniques, we found that our method was able to detect anomalous conditions that conventional approaches could not,” said Dr Mattera. “This shows that current in-process monitoring techniques are effective at identifying extreme conditions but are not suitable for preventive analysis, where the system is monitored and maintained before a defect occurs.”
In the study, traditional Statistical Process Monitoring methods struggled to detect irregularities. The results showed that the SPM approach correctly identified only 43 anomalies, while 105 were missed or misclassified as normal. By contrast, the proposed AI-based method performed far better, correctly identifying 116 anomalies and missing just 32, demonstrating its much greater accuracy in spotting potential defects.
“The results of this study are very encouraging. We’re putting significant effort into this line of research to provide industry with new tools and to demonstrate their potential impact for both manufacturers and end users,” concluded Dr Mattera. “I would like to thank the entire research team for their work over the years, as well as our industrial partners for their continued support.”
He added, “Despite the promising results, our next step is to refine the methodology and turn it into a practical tool for manufacturing. We are working on adding features such as explainability and quality index estimation, which will make it more intuitive and easier for humans to interact and collaborate with intelligent machines.”
Dr Mattera and his co-authors acknowledge this as an important step toward integrating AI into manufacturing for quality improvement. However, they note that further regulatory and legal developments will be essential for real industrial adoption. “In the meantime,” Dr Mattera concluded, “we’re pushing the research forward to make these tools ready and to help mature this field for industrial use.”
This paper ”Process monitoring of P-GMAW-based wire arc direct energy deposition of stainless steels via time-frequency domain analysis and Isolation Forest” was published in Advanced Manufacturing.
Mattera G, Manoli E, Pan Z, and Nele L. Process monitoring of P-GMAW-based wire arc direct energy deposition of stainless steels via time-frequency domain analysis and Isolation Forest. Adv. Manuf. 2025(2):0010, https://doi.org/10.55092/am20250010.
END