Optimization Learning Of Bayesian Network Structure And Parameter And Its Application In Marine Environmental Risk Assessment

The 21st century belongs to the ocean.The "Ocean Power" strategy and the "21st Century Maritime Silk Road" initiative have become the basic national policy and leading direction for China’s development.Effective access to marine environmental information,accurate prediction of changes in the marine environment,and scientific assessment of marine environmental risk are both urgent needs for marine scientific research and marine safety assurance,as well as realistic and difficult issues that need to be carried out.The incompleteness of information,the uncertainty of knowledge,the multi-source of environmental factors and the nonlinear mechanism of action in the marine environment cause great uncertainty in marine environmental security.An important part of marine security is the conduct of marine environmental risk assessment and decision support research.The core connotation of marine environmental risk assessment is the extraction,expression and modeling of uncertain information.Aiming at the fusion,reasoning and evaluation modeling of multi-source information in marine environment,this paper introduces the uncertainty artificial intelligence algorithm-Bayesian network(BN)theory,to carried out innovative exploration and research on important issues and key technologies such as marine environmental safety assurance,risk assessment and decision support.The main work and results are as follows:(1)We systematically analyzed and elaborated the uncertainty of risk,and summarized the uncertainty characteristics as original uncertainty,information uncertainty and cognitive uncertainty.Through the uncertainty analysis of marine meteorological and hydrological factors combined with the uncertain risk theory,the uncertainty of marine environmental risk was generalized as randomness,ambiguity,incompleteness,instability and inconsistency.Aiming at the uncertainty of marine environmental risk,this paper proposed the idea and approaches of BN-based risk assessment modeling technology under uncertain conditions.(2)For the BN structure learning from big data,there are problems such as large error and low efficiency in the determination of structural arcs.This paper proposed a structure learning algorithm based on global causal analysis with information flow and 0/1 optimization principle-Improved Greedy Search Algorithm(AGS).Firstly,the 0/1 optimization problem was constructed based on the information flow for global causal analysis,and the optimal initial network structure was obtained.Then,the search space was generated based on the initial structure,and the optimal structure arcs were searched by the greedy algorithm.At the same time,the arc direction was determined by the information flow,to achieve integrated learning of the BN structure.Numerical experiments show that the AGS algorithm can obtain an approximate global optimal structure better than the classical algorithm.The introduction of information flow realizes the synchronous determination of arc and arc direction,which simplifies the search procedure and makes the algorithm more reliable and efficient in accuracy and time performance.(3)For BN parameter learning in the actual evaluation application,the training sample data is not quantitative and the information is incomplete.The existing algorithm has shortages of easy converge to local optimum and slow learning for parameter learning under missing data conditions.Based on the basic theory and technical flow of genetic algorithm(GA),this paper constructed the error function to realize the error feedback between the observed information and the inference information.The GA was used to reverse the optimal probability distribution of the node,and the training of the network parameter was transformed into the optimization problem of multivariate function.The inversion technique of network parameter learning under the condition of non-quantitative data and incomplete information was proposed.Numerical simulation and experimental results verify the effectiveness,feasibility and practicability of the inversion technique.(4)For the conditional independence of the BN,it is difficult to meet the hypothesis in the actual evaluation problem.This paper uses the weight distribution between variables to improve the conditional independence hypothesis.On the basis of the naive Bayesian framework and the existing weighted BN,the subjective problems in the weight calculation method of network nodes are summarized.The traditional statistical method-grey correlation analysis is selected and improved.The weighted BN model is optimized by using the improved grey relational analysis to calculate the network weights.An optimization model for weighted BN is established.(5)Based on the improved BN,dynamic BN,cloud model and other intelligent algorithms,the evaluation and prediction of marine environmental risk in the sea area along the 21st Century Maritime Silk Road were carried out.Firstly,according to the difference of actual application conditions,different modeling methods are used for risk assessment,and two sets of marine environmental risk assessment technical processes are proposed respectively.Secondly,dynamic BN is used to model and carry out dynamic assessment and prediction of risks.The artificial intelligence technology is used to evaluate and analyze the marine environmental risk,and the risk assessment results and experimental risk zoning of 18 coastal port cities and sea areas along the South China Sea-Indian Ocean coast are given.