Research On Liquefaction Prediction Model Of Liquefiable Cargo Based On Machine Learning

Iron concentrate,a powdered form of iron ore with iron content exceeding 60%,exhibits porous characteristics and forms a stable skeletal structure.However,during maritime transportation,the cargo may experience liquefaction due to external loads,posing a significant risk to vessel safety.To prevent the occurrence of liquefaction,the International Maritime Organization(IMO)has issued the International Maritime Solid Bulk Cargoes(IMSBC)Code,which sets strict regulations and standards for the shipment and transportation of flowable bulk cargoes such as iron concentrate.Engineering experiments and testing play a crucial role in this process.Nevertheless,conventional testing methods have limitations and cannot guarantee that iron concentrate cargo will remain non-liquefied throughout transportation.Furthermore,with the continuous development of energy conservation,emission reduction,vessel intelligence,and digitization,more reliable methods are needed to ensure the safe transportation of bulk carriers.Therefore,this study establishes a series of predictive models based on numerical simulation and experimental data to forecast the liquefaction tendency of iron concentrate cargo,providing a more dependable safeguard for the secure transportation of iron concentrate.To analyze the liquefaction issue of saturated iron concentrate under load conditions,numerical simulations of saturated iron concentrate were conducted.The approach employed was based on FLAC modeling of a full-scale cargo hold.Data on pore water pressure and effective stress at the top,middle,and bottom portions of the iron concentrate were collected.The liquefaction degree of the cargo under different external excitations was assessed using the maximum excess pore pressure ratio.Additionally,statistical analysis and Pearson correlation coefficient analysis were performed to further elucidate the correlation between characteristic parameters and the maximum excess pore pressure ratio.The results indicate that,under heave motion,the median of the maximum excess pore pressure ratio induced by saturated iron concentrate concentrates towards the upper quartile,indicating that heave motion is more likely to trigger liquefaction of saturated iron concentrate.Furthermore,there is a positive correlation between frequency and acceleration and the maximum excess pore pressure ratio.The relative density exhibits a negative correlation with the maximum excess pore pressure ratio,although it is not very pronounced.In order to fully utilize the results of numerical simulations and reduce the computational time required,this study utilizes the numerical simulation results to establish regression prediction models such as Support Vector Regression(SVR),Artificial Neural Networks(ANN),and ensemble learning methods(Bagging and Ada Boost).The Bayesian algorithm is employed to optimize these models,aiming to further reduce the uncertainties caused by manual parameter tuning.The results show that the established prediction models perform excellently on both the training and testing sets,with R-squared values exceeding0.9.However,the ANN model demonstrates superior prediction accuracy,especially for larger maximum excess pore pressure ratios,when applied to new samples.This approach provides a novel perspective for predicting the liquefaction of saturated iron concentrate.To simulate the scenario where the bottom of the cargo hold in a bulk carrier contains saturated iron concentrate while the upper portion remains unsaturated,an unsaturated iron concentrate material column test was conducted.Additionally,to ensure the authenticity of the collected pore water pressure data,an experimental group and a control group were set up to analyze the consistency of their trends.Based on this,time series prediction models such as Gated Recurrent Unit(GRU),Bayesian Optimized Long Short-Term Memory(BOLSTM),and Convolutional Neural Network-Long Short-Term Memory(CNN-LSTM)were established to forecast the trend of pore water pressure and further reflect the shear strength of iron concentrate inside the cargo hold.The predictive results of the models,based on the data from this study,indicate that Mean Absolute Error(MAE)as the loss function for the prediction model performs better in capturing the trend of pore water pressure compared to Mean Squared Error(MSE)as the loss function.In the case of single-step prediction,taking MAE as the loss function and an input window size of 5 as an example,the GRU model achieves an R-squared value of 0.964 and a maximum deviation(Pa)of 7.802,improving the R-squared value and maximum deviation by 0.130 and 6.939,respectively,compared to the LSTM model.In terms of multi-step prediction,the model that combines CNN and LSTM outperforms the LSTM and GRU models in predicting the pore water pressure values for the next 5 time steps,with R-squared values improved by 0.287 and 0.034,respectively.This method demonstrates good accuracy in predicting the trend of pore water pressure and holds practical significance.