1.

INTRODUCTION

There are significant challenges faced in fabrication and then the subsequent life cycle of AM parts when seeking to ensure needed initial quality and reliability. The inspection and characterization of powder metal parts at various points in the manufacturing process has been under consideration for several decades, with much of the attention focusing on non-destructive testing (NDT) of finished or near-finished parts^1,2. Over the years some methods have been demonstrated for application at interim fabrication steps and at other interim points during manufacture^3-5. Critical to understanding NDT needs is material characterization and “allowables” (those “naturally” occurring material anomalies that are acceptable), such as some level of micro-porosity and grain variation that will not impact performance under some defined set of stressors⁶.

Providing the capability for in-situ monitoring has the advantage of not only inspection of the part to ensure that it is produced to meet the design requirements, but also to detect anomalies which occur within the build process⁷. There are several technologies that are currently deployed, including those using optical⁸ and IR imaging⁹. The advantage of using acoustic and ultrasound-based techniques is that the acoustic signatures have the potential to contain rich information about the material properties, defects and anomalies, as well as the process conditions.

2.

ACOUSTIC IN-SITU PROCESS MONITORING

Acoustic and ultrasonic-based methods have the advantage of being able to be deployed to give real-time continuous monitoring of a part during the manufacturing processes. Such techniques can be used for process monitoring with different forms of sensors¹⁰, with laser generated and detected ultrasound ^11,12 and for monitoring the acoustic emissions from cracking events⁸. Considering the successful application of acoustic-based monitoring techniques for other manufacturing processes, these techniques seem to have the potential for in-situ additive manufacturing process monitoring ^13,14 and some new systems and techniques have been developed for this purpose¹⁵. For reliable monitoring of the additive manufacturing process, it is important to identify signatures, develop metrics and the transient process-related signals in the presence of potentially high levels of time-varying noise, generated by the AM machine and processing environment. Both temporal and spectral features can potentially carry useful information that are correlated to the process and part conditions and which can be used for quality monitoring purposes.

Experimental setup and data collection

An instrumentation system together with an experimental fixture that supports piezoelectric acoustic sensors was designed to enable attachment to the control stage of a Direct Energy Deposition (DED) system. Titanium 6Al-4V powder was deposited on a steel substrate under a variety of conditions. The fixture built to support sensors for process monitoring is shown in Figure 1. The arrangement and dimensions of the specimens and the sensor locations are shown in Figure 2.

Figure 1.

Monitoring Fixture and accessories, (a) upper and lower adapter plates, build plate and mounting posts, (b) attaching the mounting plate and sensors to the upper adapter plate

Figure 2.

Direct Energy Deposition (DED) single layer deposition of Titanium 6Al-4V powder on steel build plate on 5x5 grid, (a) test grid and related dimensions, (b) location of the sensors related to the grid and on the build plate

A typical section of “RF” data record and corresponding spectrum are shown in Figure 3, which indicates that there are two main frequency bands where the majority of energy occurs. Based on this observation, the band of frequencies observed was divided into a low frequency band (<800 kHz) and a high frequency band (>800 kHz), and the dominant temporal and spectral features in each were investigated.

Figure 3.

An example of the (a) original RF acoustic signal, and (b) frequency bands segments

Temporal metrics and analysis of acoustic signatures

The signal processing employed consisted of band-pass filtering (150 kHz-2MHz) applied using a Kaiser order filter design and a convolution filter. Baseline subtraction was utilized to remove any signal drift and electronic noise using interpolated noise power spectra between intermittent baselines taken during testing (Boll), a common noise suppression technique in speech processing.

In acoustic emissions literature ‘Hits’ are defined as a signal that exceeds some predefined threshold and these are generally counted over time. They are also used to identify waveforms for further investigation and characterization. Hits were identified as waveforms with amplitudes exceeding 2 standard deviations from the mean (~95% confidence interval). The accumulation of hits over the course of a build under Normal conditions in depicted in Fig. 4. It can be observed that the hit total rises quickly during the build, and settles after build completion (~2500 waveforms). Hits have been associated in previous work with formation of material defects including porosity and cracking¹⁶ and this can be used as a material state indicator.

Figure 4.

RMS noise levels for a normal build condition showing an underlying process noise and interspersed hits. An accumulated number of the hits over the course of the build is also shown as an indicator of material state

Excluding waveforms identified as hits shows a marked decrease in RMS noise after build completion in Figure 5. Also displayed is the Baseline level RMS noise level which the Normal build condition approaches upon build completion.

Figure 5.

RMS noise values for the same Normal build condition depicted in Fig. 4 with waveforms associated with Hits removed to isolate process noise. A marked reduction in RMS noise levels is seen at approximately 2500 Waveforms coinciding with build completion and a return to Baseline RMS noise levels

Accumulating temporal metrics and fitting with normal distributions, the scatter in the central tendency (mean) and standard deviation appear to follow consistent trends as seen in Figure. 6.

Figure 6.

Clustering based on build conditions show consistency between builds of the same condition (same markers) and separation with other conditions (unique markers). Central tendencies of a fitted normal distribution to RMS noise level distributions tend to be elevated above baseline for all conditions, and with stronger deviations for builds entailing material deposition.

Spectral metrics and analysis of acoustic signatures

The use of frequency domain spectral features has proved to be particularly useful when there are a variety of different noise generation mechanisms in a system, such as in manufacturing machinery¹⁷, processing systems including boilers and heat exchangers, and in turbo-machinery fault diagnosis¹⁸. If defined appropriately, the frequency-related features and signatures for the acoustic signal are very effective in terms of “event” or source extraction and their use for discrimination and classification purposes.

The method used for data analysis used here is based on feature extraction from the frequency response of the acoustic signature signals. Defined features in the spectral domain are listed in Table 1.

Table 1.

Spectral features type and abbreviation

Clustering analysis is based on plotting the identified features in pairs and triples (2D and 3D respectively) to study the separation of variables and classification of process conditions for different build settings. The Clustering results for data from an acoustic piezoelectric sensor using centroid frequency (Cf) and centroid amplitude (CA) of spectral data obtained using the Fourier Transform in low frequency and high frequency bands are shown in Figure 7. The classification of process conditions using three frequency domain features; centroid frequency (Cf) and centroid amplitude (CA), and peak amplitude of frequency spectrum (PA) of spectral data obtained using the Fourier transform for a single sensor at low and high frequency bands is shown in Figure. 8.

Figure 7.

Clustering results for data from a single sensor using centroid frequency (Cf) and centroid amplitude (CA) of spectral data obtained using the Fourier Transform in (a) low frequency and (b) high frequency bands

Figure 8.

Classification of process conditions using three frequency domain features; peak amplitude of frequency spectrum (PA), centroid frequency (Cf) and centroid amplitude (CA) of spectral data obtained using the Fourier Transform for a single sensor at (a) low frequency and (b) high frequency bands

Distinct separations of clusters are seen in Figures 7 and 8 to correlate with the data from different process conditions. Combination of acoustic signatures in the frequency domain can provide closer and more effective data clustering for different process conditions.

3.

CONCLUSIONS

A proof-of-concept study has investigated new approaches to process monitoring for additive manufacturing based on acoustic signatures. Various alternatives for signal processing, pattern recognition, and classification methods were applied to acoustic signals generated by an additive manufacturing process. It has been shown that acoustic signal characteristics can be used to classify process and system conditions. The acoustic signals were collected during the Direct Energy Deposition (DED) additive manufacturing process operated under different process conditions. A novel application of signal processing tools is used for the identification and use of metrics based on frequency spectral features in acoustic signals for the purpose of in-situ monitoring and characterization of conditions in an additive manufacturing process. A spectral feature-based clustering method was implemented to analyse the acoustic signals. Clustering plots for metrics in 2 and 3-D were used to facilitate the visualization of the groupings and condition discrimination. It is demonstrated that a passive acoustic monitoring approach and use of signal processing algorithms is effective at giving metrics that achieve clustering and separation of conditions based on multiple spectral features extracted from the original test data, and that these metrics do correlate with different AM system conditions.

Classification of different DED additive manufacturing process conditions exhibit successful clustering of large data sets. Evaluation of the identified features confirmed the consistency in process monitoring and data collection by all sensors, different locations on the build plate, and various process conditions. Results show that this novel approach using acoustic signal analysis can provide metrics based on acoustic signals (signatures) generated by the AM process, and classification of the signatures can be correlated with different process conditions.

Monitoring of the manufacturing process using acoustic signatures would appear to have the potential to give data which can help enable early detection of off-normal conditions, generation of faults in the process, and can be used for process optimization and control.