Many experiments are usually needed to quantify probabilistic fatigue behavior in metals. Previous attempts used separate artificial neural network (ANN) to calculate different probabilistic ranges which can be computationally demanding for building probabilistic fatigue constant life diagram (CLD). Alternatively, we propose using probabilistic neural network (PNNs) which can capture data distribution parameters. The resulted model is generative and can quantify aleatoric uncertainty using a single network. Two tests are presented. The first captures the fatigue life aleatoric uncertainty for P355NL1 steel and successfully builds a probabilistic fatigue CLD. The resulted network is not only more efficient but also provides higher accuracy compared with ANN. To assess fatigue, the second test examines vibrations of a pipework assembly. The proposed methodology quantifies the nonlinear relation between the vibration velocity and the equivalent stress and successfully reflects measurements uncertainties in fatigue assessment. The proposed methodology is published in opensource format (https://github.com/MShadiNashed/probabilistic-machine-learning-for-fatigue-data).

The modelling of fatigue using machine learning (ML) has been gaining traction in the engineering community. Among ML techniques, the use of probabilistic neural networks (PNNs) has recently emerged as a candidate for modelling fatigue applications. In this paper, we use PNNs with nonconstant variance to model fatigue. We present two case studies to demonstrate the developed approach. First, we model the fatigue life of cover-plated beams under constant amplitude loading, and then we model the relationship between random vibration velocity and equivalent stress in process pipework. The two case studies demonstrate that PNNs with nonconstant variance can model the distribution of the data while also considering the variability of both distribution parameters (mean and standard deviation). This shows the potential of PNNs with nonconstant variance in modelling fatigue applications. All the data and code used in this paper are openly available.

We present a novel method for real-time fault classification using the time history of acoustic emissions (AEs) recorded from a lab-scale gas turbine operating under normal and faulty conditions across multiple turbine speeds. Time-frequency features are extracted using the continuous wavelet transform, and for each signal, the root mean square (RMS) and kurtosis are calculated. We employ a color mapping technique to combine the time-frequency and statistical features into a single red–green–blue (RGB) image. The red channel is mapped to the time-frequency data, whereas the green and blue channels are mapped to the RMS and kurtosis, respectively. Subsequently, a deep convolutional neural network is trained on the generated images to classify the gas turbine condition. We show that the proposed model can form an online monitoring system using AEs to classify multiple running conditions at various turbine speeds. The methodology not only achieves real-time classification of faults but also minimizes the human intervention in identifying these faults. The datasets and codes used in this paper will be openly available.