预测特征误差映射及其在多基地水下目标识别中的应用

针对水下多基地目标识别问题,提出了基于特征预测和误差映射的多基地融合识别算法。推导并简化了基于贝叶斯公式的多基地目标识别条件概率公式,利用BP神经网络对最后一个节点的特征向量进行预测,并计算得到预测值与实际值误差的概率密度,将其与前面每个节点的条件概率累乘,以得到目标识别的条件概率。将利用特征预测计算条件概率的方法从单个节点推广到多个节点上,同时针对误差概率分布模型不准确的问题,提出了利用混合高斯分布模型代替单高斯概率分布模型的改进方法。对每个目标重复此过程,取结果最大值对应的目标类别为最后的识别结果.在消声水池开展多基地模拟实验,对四类目标进行识别,在一定声呐节点数目及信噪比条件下,与单基地声呐相比,多基地目标融合识别得到的识别正确率最大可提高40%,采用改进方法以后,识别正确率得到进一步提高。

Predicted feature error mapping and its application in multi-static underwater target recognition

The method based on feature prediction and error mapping is used to conduct multi-static target identification. The target identification conditional probability is obtained through the Bayes rule. And the formula equation is simplified. The prediction feature of the last receiver is obtained by BP(Back Propagation) neural network, and the prediction error probabilistic is calculated using the Gaussian mixture model. The improved scheme expands the method using feature prediction and error mapping to calculate the conditional probability of multi sonar nodes and replaces the single Gaussian model with the Gaussian mixture model to solve the problem of inaccuracy of error probability distribution model. Then the multiplied result by the identification probability of each former receiver is used to obtain the conditional probability of the target type. The procedure is repeated for every target, and the target type corresponding to the maximum probability is the type of target to be identified. Multi-static simulation experiment is conducted in anechoic tank, and the above method is employed to identify four types of targets. In the condition of certain sonar numbers and signal-to-noise ratio, compared with mono-static sonar system, after employing the multi-static fusion recognition method, the identification rate of multi-static sonar system is increased by 40% and the identification rate is increased more after the method is improved.