A data logging sampling frequency invariant ANN approach for generalised random fatigue loading damage analysis

• Introduced a sampling frequency invariant approach for random loading fatigue analysis. • Approach implies missing sampling frequency information will not impede the analysis of any time domain random fatigue loading data using frequency domain and ANN methods. • Remarkable agreement is obtained between ANN and time domain methods and also with SAE variable amplitude cumulative damage experimental results within 99.7% scatter confidence level. • A quantitative analysis of the computer execution times using time and frequency domains is presented. • Some Input data generated for the work, limited by the space allowed, is available on GitHub. This paper presents a novel data logging sampling frequency invariant ANN model for generalized random loading fatigue damage analysis. The approach, which is original, enables an ANN trained using one sampling frequency to be universally applicable to data logged using any other Nyquist appropriate sampling frequency. It also covers situations where data logging sampling frequency information is missing altogether. The approach is based on using the Parseval’s time and frequency domain energy equivalence theorem and, secondly, appropriate frequency power normalisation for each spectral-related parameter required as input for the ANN model. This leads to the elimination of dependency on the original data logging sampling frequency. The ANN model was trained on simulated random loading data subjected to physical constraints representing feasible loads and limits of alloy material fatigue properties. The model was tested on an extensive random fatigue loading experimental data produced by a consortium led by the Society of Automotive Engineers (SAE) for cumulative damage analysis. The Pearson correlation coefficient, R , between the time and the ANN model results was > 0.999 with the latter method being much faster to run, by up to a factor of 20 or greater. The results of the ANN and the experimental testing that were within a factor of 3 scatter band was about 82 % for the best performing method of accounting for the effect of the stress concentration in the SAE specimens. This performance for the ANN model is considered very good given the significant editing of the original time domain signal.