Research On Survey Technique For The Corrosion Protection Status Of Submarine Pipelines

As the development of the offshore petroleum and gas exploitative engineering, more and more submarine pipelines were used broadly. The invalidation of the corrosion protection is the main reason for the accident of submarine pipeline. Therefore, it is very important to inspect the status of submarine pipeline protection.The status of submarine pipeline protection can be described with several factors, including the average coat defection of sacrificial anode, the current of sacrificial anode, remaining weight of sacrificial anode and etc. The inspection data is more accurate through the measurement of potential difference. Meanwhile, the simulated calculation for the integrated model of submarine pipeline under cathodic protection system is made in order to build the database,'infection factors database of electric field under cathodic protection system'. Based on the database, the non-linear relationship between potential difference and those determined influence factors of the status of submarine pipeline cathodic system can be mapped by neural network training. Through those researches, the inspection for status of the submarine pipeline cathodic system can be realized by measuring potential difference surrounding the submarine pipeline.Through the analysis of actual environment of submarine pipeline, 720 simulation calculated models for the submarine pipeline cathodic system are built and the database, 'infection factors database of electric field under cathodic protection system', was built after all of those models were calculated. After analysis of the relationship between 'input parameters' and 'output parameters', the feasibility of applying neural network algorithm to calculating the potential and electric current density of the submarine pipeline is proved correct.Being a non-linear training algorithm, the algorithm of Back-Propagation has many merits and is used in large range. Admittedly, there are also some limitations. In order to mend those limitations, four neural network training algorithms, Levenberg-Marquardt algorithm, Momentum Back-Propagation algorithm, Advanced Stop Back-Propagation algorithm, Bayesian Regularization algorithm, are applied in our research. .According to the difference of input parameter, four sorts of neural network are designed. A series of neural network structures, which are created through the combination of different hide layer number and different neural cell number, are trained by different training algorithm and training data. The optimized neural...