This study aims to develop a robust method for analyzing the attitude status of Geostationary Earth Orbit （GEO） satellites based on photometric data. Due to their high orbital altitude， ground-based telescopes cannot resolve imagery of GEO satellites， making photometric analysis a crucial technique for attitude monitoring. Existing methods primarily rely on time-series analysis but often lack comprehensive multi-dimensional feature extraction and validation. To address this gap， we propose constructing a photometric feature set that integrates time-domain， frequency-domain， and sample entropy characteristics. Furthermore， this study evaluates the classification performance of these features using both traditional machine learning algorithms and deep neural networks. The primary objective is to establish an effective and interpretable feature system that can reliably detect satellite attitude anomalies， thereby providing technical support for practical space situational awareness applications.To achieve this， the research methodology consists of four main components. First， we collected and preprocessed measured photometric data of GEO targets， performing state labeling to create a validated dataset containing both normal and anomalous satellite states. Second， we analyzed the optical scattering characteristics of high Earth orbit space targets to establish a theoretical foundation for feature selection. Based on this analysis， we constructed a multi-dimensional feature set comprising： （1） time-domain features （standard deviation/variance and fluctuation coefficient）， which capture statistical characteristics and overall variability of light curves； （2） frequency-domain features （average amplitude， center frequency， and frequency variance）， derived from power spectral density analysis to quantify periodic components and energy distribution patterns； and （3） sample entropy， a nonlinear metric that measures time series complexity and randomness. Third， we implemented five classification algorithms： three traditional machine learning methods （Support Vector Machine， Random Forest， and Decision Tree） and two deep neural network approaches （Bidirectional Long Short-Term Memory and Fully Connected Neural Network）. Finally， we conducted controlled experiments to evaluate feature effectiveness through average classification accuracy， comparing single-feature performance， time-domain versus frequency-domain features， and the complete multi-dimensional feature set.Experimental results revealed several key findings regarding feature effectiveness. Frequency-domain features demonstrated superior discriminative power， with average amplitude achieving the highest single-feature classification accuracy of approximately 79.54%， significantly outperforming other individual features. The top three performing features were all frequency-domain characteristics， whereas time-domain features and sample entropy showed notably lower accuracy. Sample entropy performed poorly in this task with average accuracy around 60.25%， approaching random classification levels， likely due to measurement noise in normal satellite data increasing entropy values and obscuring distinctions. When comparing feature categories， frequency-domain features consistently outperformed time-domain features across all algorithms， with accuracy advantages of approximately 16 percentage points： Random Forest achieved 90.03% with frequency-domain features versus 73.40% with time-domain features； FCNN reached 86.04% versus 65.39%. Most importantly， the multi-dimensional feature set integrating time-domain， frequency-domain， and sample entropy characteristics achieved the highest classification performance across all algorithms： Random Forest reached 94.16%， Decision Tree achieved 90.58%， FCNN attained 89.71%， SVM reached 82.31%， and Bi-LSTM achieved 80.43%. This represents improvements of 4-6 percentage points over frequency-domain features alone， demonstrating the complementary nature of multi-dimensional features.These findings demonstrate that frequency-domain characteristics serve as the most reliable indicators for detecting GEO satellite attitude anomalies， with average amplitude identified as the optimal single-feature discriminator. The superior performance of frequency-domain features stems from their direct physical connection to satellite dynamics： normal attitude-stabilized satellites exhibit energy concentrated near zero frequency， while anomalous tumbling motion introduces periodic components that disperse spectral energy—a phenomenon that time-domain statistics cannot adequately capture. Multi-dimensional feature fusion significantly enhances classification performance by integrating complementary information： time-domain features capture overall statistical fluctuations， frequency-domain features quantify periodic structural changes， and nonlinear entropy measures signal complexity. These findings provide practical guidance for space object monitoring： prioritizing frequency-domain features while incorporating complementary dimensions offers the most effective approach for photometric-based satellite state diagnosis. Future work should focus on expanding anomalous state sample diversity and exploring alternative complexity metrics better suited to photometric data characteristics.