随着各国航天活动的增多, 空间目标的数量和种类不断增加, 对空间目标进行编目识别是各国空间目标监视领域的重要研究内容。 对空间目标进行识别, 主要是为了获得其表面材质、 姿态、 形状、 关键载荷等信息, 而表面材质信息的获取是开展目标光学特性及状态认知研究的基础。 搭建空间目标表面材质多色测光测量系统, 整套系统部署在光学暗室内, 以减少杂散光对测量结果的影响。 光源采用太阳模拟器, 光谱等级A级; 探测器采用美国ASD公司生产的FieldSpec4地物光谱仪, 波长范围350～2 500 nm, 光谱分辨率1 nm, 光纤置于电控转台上, 能对待测样片实现不同观测几何下的测量。 利用Johnson-Cousins UBVRI五色分光系统对8种常用表面材质(砷化镓、 氧化铝、 氧化聚酰亚胺薄膜、 黑漆、 环氧漆、 镀铝聚酰亚胺薄膜、 钛青蓝漆、 白漆) 在不同观测几何条件下的10种色指数数据进行实验测量, 每种色指数分别测得30组实验数据。 采用传统的1-sigma不确定框方法(即对于给定材质的若干组实验数据, 计算其每种色指数的平均值和标准差, 以平均值为中心, 以标准差的两倍为边长画出色指数不确定框) , 在最理想的识别情况下, 通过R-I和B-R色指数不确定框能对砷化镓、 氧化铝、 氧化聚酰亚胺薄膜、 钛青蓝漆四种材质进行识别; 利用B-V和B-R色指数不确定框可以将环氧漆、 白漆识别出来, 剩余两种材质黑漆和镀铝聚酰亚胺薄膜无法通过以上色指数进行识别。 但是1-sigma不确定框方法存在两个主要问题: 一是需要知道待测材质对特定波段敏感的先验信息, 来确定所用的色指数类型; 二是识别率容易受测试样本数量的影响, 可靠性差。 超限学习机算法是一种利用随机化隐层节点和和最小二乘求解方式进行训练的机器学习算法, 具备学习效率快, 泛化性能好, 不容易陷入局部最优解等优势, 被广泛应用于对数据的分类和回归分析中。 因此引入超限学习机算法, 将色指数数据按照2∶1的比例随机分为训练样本和测试样本, 共进行三次随机试验。 在训练样本中, 对每种材质按照1∶8的顺序进行编号, 即编号1∶8的测试样本分别有20个, 分别包含10种色指数数据; 在测试样本中同样对其按照已知归属材质对应编号。 采用决定系数和计算时间作为判断ELM算法准确性和实时性的判断指标。 结果表明: 无论是对单一材质进行识别, 还是对所有测试材质样本, 训练样本决定系数在0.98以上, 测试样本决定系数在0.96以上, 每次试验中最多有3组色指数数据无法识别; 所需总时间最长在0.07 s内完成, 甚至可以达到0.002 s, 识别效率和可靠性远高于传统的1-sigma不确定框法, 表明ELM算法能对空间目标常用材质进行准确快速识别。 相关研究可为空间非合作目标的外形、 姿态等状态信息反演提供技术支持。

With the increase of space activities in various countries, the number and variety of space objects also gradually increased. How to identify and catalog the space object is a critical research issue in the field of space object surveillance for different countries. The research on non-cooperative space object mainly aims to get the information like surface materials, attitude, shape and critical payload information. And the acquisition of surface materials information is the basis for researching space object optical characteristics as well as state recognition. A multi-color photometric measurement system for space object’s surface materials is set up. To reduce the influence of stray light on the measurement results, the entire system is deployed in an optical darkroom. The light source adopts a solar simulator with spectral grade level A. The detector uses a FieldSpec4 spectrometer manufactured by the ASD company in America. The wavelength range is 350~2 500 nm, and the spectral resolution is 1 nm. The spectrometer’s optical fiber is located on the electrically controlled turntable, which can be able to simulate different observation geometry to obtain various data for the same sample. By using Johnson-Cousins UBVRI five-color spectroscopy system, ten kinds of color-index data of eight common surface materials (e.g., GaAs, anodic Al, anodic Kapton, black paint, epoxy paint, aluminized Kapton, titanium blue paint, white paint) under different observation geometry conditions are measured. And every color-index data includes 30 groups of experimental data. Through the traditional 1-sigma uncertainty box method (namely, for a given material with several groups of experimental data, calculating the mean value and standard deviation for each kind of color-index. Then drawing color-index uncertainty box, based on the mean value as rectangle center and twice of standard deviation as the side length), in the ideal identification situation, four kinds of materials (GaAs, anodic Al, anodic Kapton, titanium blue paint) can be identified through the R-I and B-R color index uncertainty box. Two kinds of materials(epoxy paint, white paint) can be determined through the B-V and B-R color index uncertainty box. That above color-index cannot identify the remaining two materials (black paint and aluminized Kapton); however, there are two main problems in using 1-sigma uncertainty box method. The first one is that it is necessary to know the prior information about which band that the test materials are sensitive to, so as to determine which kind of color-index to be used. The other problem is that the identification rate is easily affected by the number of test samples and has poor reliability. The extreme learning machine (ELM) is a kind of machine learning algorithm that uses randomized hidden layer nodes and least-squares method to train data. The algorithm has the advantages of learning efficiency, good generalization performance and not easily falling into optimal local solution. It is widely used in the data classification and regression analysis. Therefore, the ELM algorithm is introduced to solve the problem. Color-index data are randomly divided into a training dataset and testing dataset according to the proportion of 2∶1. A total of three random experiments are carried out. Each kind of material is numbered in the order of 1∶8 in the training dataset, namely, each number from 1 to 8 has 20 groups color-index data respectively, and each group includes ten kinds of color-index data. As for the training dataset, the same number is assigned to the material according to the known attribution materials type. Regard determination coefficient and calculating time as the judgment indicators to judge the accuracy and real-time capability of ELM algorithm. The results show that: whether identify the single kind of material or all testing dataset, the determination coefficient of training dataset is all above 0.98 and determination coefficient of the testing dataset is above 0.96, meaning that at most three groups color-index data cannot be identified in each experiment. In the aspect of calculating time, the total time can be completed within 0.07 s, even up to 0.002 s. The identification efficiency and reliability are much higher than the traditional 1-sigma uncertainty box method, which shows that the ELM algorithm can accurately and quickly identify space object’s surface materials. Relevant research can provide technical support for state information inversions such as shape, attitude and critical payloads of space objects.