Improvement of accuracy for egg dynamic weighing using artificial neural network

Abstract. Weight is an important factor for egg grading in modern egg processing equipment. The oscillatory underdamped weighing signal from strain gauge load cell is usually contaminated by various vibration disturbances which predominantly arise from the rolling egg and dynamic mechanical parts. Usually, the arithmetic mean of the values in the plateau of denoised load cell signal was used as the final mass estimation. However, the stochastic disturbances were time-varying and reflected in time domain of load cell signal in a form of strong amplitude fluctuations which seriously distorted the waveform and hence degraded the weighing accuracy. In this research, the procedure of mass measurement was regarded as the function fitting problem, and a multilayer perceptron(MLP), which was a two-layer feedforward artificial neural network, was employed to improve the egg dynamic weighing accuracy. Five features, including three summary statistics (arithmetic mean, median and 25% trimmed mean), and the maximum and minimum value in the analysis interval of the denoised load cell signal were extracted as inputs to the MLP. An arc-shaped rail based egg dynamic weighing apparatus was used for egg dynamic weighing experiments at three different processing speed. Comparing with the commonly used low-pass filtering and a subsequent average-based mass estimation method (AME), such as arithmetic mean, median, and trimmed mean based mass estimation, the experimental results showed that the developed MLP could improve the weighing accuracy effectively and apply to different processing speed.