Typically, VAM signals are evaluated in the literature by means of estimating a single representative value from the Fourier-transformed acoustic spectra, i.e. a “modulation intensity coefficient” (R), the MI, and more. The calculation of these representative values is similar, but all values have an emphasis on the first sideband in common, which mainly results from an amplitude modulation of the original signal. The ultimate goal is to be able to relate their value to a specific damage state and thus enable a residual lifetime prognosis. Above values are useful
when large changes in the structural response occur, resulting from large damages in the sample. However, signal modulation is a superposition of amplitude, frequency, and phase modulation, which are difficult to distinguish and most likely result from the superposition of response signals from different damage types.
When machine learning methods are applied to the VAM signal evaluation, all sidebands and thus the inclusion of the superposition of all modulations may be more rigorously integrated. For this purpose artificial neural networks (ANN) were trained in a supervised learning approach to identify patterns in the vibroacoustic frequency spectrum. 
The identification of specific patterns, that are not directly visible to the human eye or intuition, makes it possible to predict the physical state of the sample. In parallel to the development of an ANN, allowing a robust prediction of the physical state of a sample—an example is given below—an analytical solution to predict the lifetime of FRP samples and the corresponding damage state is pursued with a data-driven approach (cf. Figure). A clear logarithmic trend of the MI correlating with the “damage” of the sample could be observed. Furthermore, the logarithmic trend may be
related to the bending modulus of the specimen and the maximum load applied by the cyclic testing machine, so that a universal description of the damage condition from the MI is possible for GFRP specimens with e.g. different fiber orientations and lay-ups. The application of machine learning is hence not yet essential but may become more relevant when the data set grows and different fibre type, lay-ups, geometries and testing amplitudes are included.
Furthermore, the detection of adhesive joint defects in bonded structures was investigated by analyzing the vibroacoustic signals using an ANN. Currently, there is no method to detect the occurrence of weak-bonds nondestructively. (Ultrasonic) pulse-echo techniques are able to detect an artificially introduced film in the bond, but surface contaminations—e.g. silicone, which even in small amounts drastically reduce the bond strength—are typically invisible (see “Pristine” and “Release Agent” in the left panel of Figure 1). To investigate this at a more fundamental coupon level, single-lap shear specimens featuring a bond size of 25mmx25mm were prepared. An artificial disbond was created in some of the samples by introducing a PTFE sheet in the joint. Weak-bonds, on the other hand, have been artificially manufactured by applying circular spots of release agent on the substrates. The PTFE sheet is clearly detectable as a red area in the pulse-echo C-scan in Figure 1. 
Furthermore, PTFE samples illustrate a significant change of the MI and are easy to determine as well with a VAM analysis. Artificial weak bonds, on the other hand, do not show a significant difference from the original sample in the C-scan, nor are they detectable by changes in the parameters MI, b or R estimated from the VAM signal. However, the application of a deep neural network trained on several sidebands made a detection possible. A trained ANN on 80% of the specimens (with a total of 38) is able to differentiate between disbond, weak-bond or pristine sample with an accuracy of 90.5%. The accuracy of the prediction increased to 93% when only a distinction between weak bonds or pristine samples was required. This result emphasises the striking relevance of evaluating more than just the first sidebands in order to facilitate a more precise statement about the damage state and underlines the potential of enhancing VAM with the application of artificial neural networks. A publication summarizing the findings on weak bond detection is currently in preparation. To apply VAM to structural components, a thorough analysis of the initial state of the structure must be performed to define the initial state for VAM. Each measurement is then compared to detect anomalies or changes in the signal. For this purpose, unsupervised machine learning methods are best suited to evaluate cumulated data from sensor networks.