3.5.1 研究目的
The project presents a method of passive localization for acoustic source in shallow water based on Ray acoustic theory. Ray acoustic theory is used to establish underwater coustic channel and to analyze physical quantities, arrival time and arrival angle, received by sensors. Then, mutual relationship of rays between sources and receivers is determined by using artificial neural network (ANN) for source localization.
3.5.2 研究要點
There are two approaches for the inverse computation of source localization. One is to measure the arrival time and arrival angle by using a hydrophone. The other is to evaluate sets of arrival angles by using array hydrophones. Both these inverse
approaches are established by ANN training to compute sound source location. In this study, the inverse computational method is not only used in shallow water, but it is also verified by using a non-trained source to proof its accuracy and reliability
In this present study, Ray theory is used to express propagation of sound wave in shallow water. The physical quantities such as: arrival time and arrival angle, are evaluated by use of ray theory to establish data basis and are used as inputs in artificial neural network. The network with input of ray physical quantities and output of source location is obtained. Therefore, when arrival time and arrival angle are measured or simulated as input, inversion of source location can be achieved by backward propagation network (BPN). The flowchart of the entire simulation and the practical experiment is shown as figure 3-11
Figure 3-11 Flowchart of simulation and experiment structure for source localization
3.5.3 模擬成果分析 Test cases
The range distance error shown in Eq. 11 defines as absolute value of actual sound source distance minus inverse computed distance from ANN divided by actual sound source distance in the horizontal direction. The depth distance error defines as absolute value of actual sound source distance minus inverse computed distance from ANN divided by actual sound source distance in vertical direction.
Case 1
The source is 18765m apart in range, and the hydrophone is 40m in depth. The acoustic rays that the sensor receives are analyzed to find the highest pick value of arrival angle and travel time of rays, and then replace the results as inputs of artificial neural network then the outputs (range and depth) are calculated by the network. There is 1m difference of the inverse calculation in the range distance, which is an error of 0.0053%. A difference of 2.85m of the inverse calculation in the depth of source is an error of 2.85%.
Case 2
Three sources are located at 26670m, 27002m, and 26000m apart in range respectively, and three hydrophone receivers are located at 30m, 45m, and 100m in depth respectively. The acoustic rays that the hydrophone receive are analyzed to find the highest pick value of arrival angles from VLA 16 pairs of rays and then replace the results as inputs of artificial neural network then the outputs are calculated by the network. There is 33m difference of the inverse calculation in the 26670m range distance, an error of 0.124%, and a difference of 0.329m of the inverse calculation in the depth of source is an error of 0.165%.
There is 55m difference of the inverse calculation in the 27002m range distance, an error of 0.204%, and a difference of 4.569m of the inverse calculation in the depth of source is an error of 2.285%. There is 27m difference of the inverse calculation in the 26000m range distance, an error of 0.104%, and a difference of 0.567m of the inverse calculation in the depth of source is an error of 0.284%.
Figure 3-12 presents the variation of the training error with respect to the number of hydrophone receivers. That result indicates the more numbers of hydrophone receivers for arrival angles increase, the fewer errors occur. The x axis means the number of hydrophone receiver. Each hydrophone receiver has 420 arrival angles, and the average value is calculated as the arrival angle at the hydrophone location. Those physical quantities are used as input data of BPN. The y axis presents the mean square error of training rate by ANN.
Figure 3-12 Comparison of ANN training error (Case 2)
Figure 3-13 shows the horizontal and vertical distance error with respect to
number of hydrophone receivers. The x axis means the number of hydrophone receiver.
Each hydrophone receiver has 420 arrival angles, and the average value is calculated to represent the arrival angle at the hydrophone location. Those physical quantities are used as input data of BPN. The y axis presents the distance error rate of inverse calculation by ANN. The black color means the range error rate and the gray color means the depth error rate. The horizontal distance error is less than 100m and vertical distance error is less than 10m. The distance error rates below 5% are acceptable.
Figure 3-13 Comparison of distance error (Case 2) 3.5.4 結論
This research is to develop a new inversion procedure for source localization in shallow water. In this research, the physical quantities (such as: arrival time and arrival angle) of acoustic ray are evaluated by ray theory. And these are cooperated with neural networks to calculate the acoustic source position. Main conclusions are generalized as the followings:
3.5.4.1 Two inverse approaches:
In the present study, two inverse approaches for the source of localization are present. One is to measure the arrival time and arrival angle by using a hydrophone.
The other is to evaluate sets of arrival angles by using array hydrophones. These inverse approaches by ANN can be applied to calculate the ocean environment parameters.
3.5.4.2 The effects of the discharged angle:
In this article, we launch the acoustic ray by controlling the discharged angle in the scope of ±11°. From ray number - time distributed chart, it can prove that the discharged angle in the scope of ±11° is suitable and feasible. Therefore, it will reduce many unnecessary operationsand savecomputer’sload ifwechoosethescopeofray discharged angle appropriately. Besides, the results indicate that control of the
discharged angle reduces the influence of multiple paths in the sea effectively from the relationship of the discharged angle By controlling the discharged angle, the threshold of number of ray reflections will filter out the rays whose energy were dissipated because of too many reflections.
3.5.4.3. The effect of grid size of acoustic source:
The initial simulated grid is divided into one division per 100 meters in range (away from the receiver 18 to 20 kilometers for case 1 and 25.26 to 27.26 kilometers for case 2) and one division per 10 meters in depth (10 to 120 meters in depth for case 1 and 2.3 to 192.3 meters in depth for case 2). The results indicate good inverse
mechanism to calculate the distance between source and receiver in range and in depth.
Also, it proves that the more divided grids, the more precisely we can construct the net.
Therefore, it is concluded that if we had enough operating time to set up more
extensive grids, then the inverse calculating mechanism of source position can be more accurate.