Abstract: Cadmium is considered one of the most harmful heavy metals because of its wide range of dangerous contamination, high toxicity, and easy invasion. Long-term intake of excessive cadmium can cause many diseases including cancers. Cadmium content prediction in the wheat processing chain is of great practical importance for developing countermeasures to reduce its hazards. In this paper, we proposed a deep learning prediction model using the regularization method to address the problem that the data of cadmium content in the wheat processing chain contain strong nonlinear and random noises, which leads to poor fitness of the traditional model. Firstly, a gated recurrent unit (GRU) was used to build the deep learning prediction model. Secondly, the loss function of the model was modified using the regularization method to reduce the impact of noise on the prediction performance of the model by adding a noise penalty term to fade out the noise fit of the model during training. Finally, a Bayesian optimization method was used to select the hyperparameters to ensure that the model could accurately predict the cadmium content at each stage of the wheat processing chain. The prediction results show that flour made from wheat grains with a cadmium content less than 0.1 mg/kg can basically meet the requirements of the national standard (GB 2762-2017).

Key words: wheat processing chain; cadmium; predictive models; gated recurrent unit; Bayesian optimization