Predicting trucking accidents with truck drivers’ safety climate perception: An in-depth evaluation of the pretrain-then-finetune approach

There is a rising interest in using AI-powered safety analytics to predict accident outcomes. However, companies face the challenge of developing accurate safety analytics models. One possible solution is to use a pretrain-then-finetune parameter-transfer learning approach for a “target company” to utilize knowledge generated from the data of other “source companies”. However, transfer learning is uncommon in safety analytics due to reasons such as lack of publicly available large-scale pre-training data and pre-trained models, differences between the source and target companies, and lack of guidelines. To fill the above gaps, we conducted experiments to study the effectiveness of transfer learning in the context of using truck drivers’ safety climate data for predicting trucking accidents. To facilitate the experiments, we developed SafeNet, a deep neural network algorithm for classifying accident outcomes. The safety climate survey data are from seven trucking companies with different sample sizes. Three new evaluation indicators are proposed to evaluate the difference between finetuned models and models trained from scratch. The study shows that transfer learning is not effective in about 20% of the cases that used pretrained models trained on one source company’s data. Instead of sample size and data diversity, the study found that accuracy of the pretrained model is more important. The trucking industry may, thus, consider developing different pretrained models for different types of companies. To promote the adoption of transfer learning in safety analytics, we make our code and pretrained models publicly available at https://github.com/NUS-DBE/Pretrain-Finetune-safety-climate.