✨ 长期致力于近红外光谱、绿茶、特征光谱波长、光谱感知节点、入射光学系统研究工作，擅长数据搜集与处理、建模仿真、程序编写、仿真设计。
✅ 专业定制毕设、代码
✅如需沟通交流，点击《获取方式》
（1）绿茶产地溯源与内在成分定量NIR分析模型：
采集山东日照绿茶和崂山绿茶样本各120个，使用爱万提斯光纤光谱仪（1000-2500nm）采集漫反射光谱。光谱预处理采用Savitzky-Golay平滑（窗口15，阶次2）和标准正态变量变换。产地鉴别采用PLS-DA模型，潜变量数为8，预测集识别率100%。首次提出特征波长-判别分析方法，提取6个特征波长后识别率仍达98%以上。时令鉴别建立PLS定量模型，以生产日期的年积日为伪标签，校正集Rcal=0.952，预测集RMSEP=19.97天。有效成分定量中，茶多酚预测相关系数达到0.943，儿茶素0.705。发现物质含量与预测精度呈正相关，含量高于10%的成分预测R>0.9。低分辨率光谱仪（分辨率15nm）与高分辨率（6nm）对比，分类准确率仅下降2.3%，证明低分辨NIR可用于现场检测。
（2）移动窗口-BP神经网络与诱变遗传算法特征波长提取：
MW-BPANN算法将光谱划分为32个窗口（宽度30nm），每个窗口训练BP神经网络，选择分类精度最高的窗口中心波长。诱变遗传算法模拟生物诱变过程，在每一代对染色体实施随机诱变（突变率0.15），五步迭代后从156个波长中筛选出11个特征波长，包括1472.76nm（蛋白质N-H伸缩）和1540.29nm（O-H弯曲）等。诱变GA相比传统GA收敛速度提升40%，特征波长数量减少70%。提取特征波长后SVM模型预测识别率从91.67%升至100%。开发了ARCO-NIR软件，集成预处理、特征提取和建模功能，采用GUIDE开发，支持批量处理。
（3）光谱感知节点入射光学系统与微型化设计：
以上海技物所光谱组件为核心，设计透射式入射光学系统，视场角1.8-14.8°可调，探测距离50m。光学系统由柱面镜（焦距50mm）和可变光阑构成，光斑整形后进入分光组件。波长范围1355-1565nm，分辨率7nm，波长准确性2nm。节点硬件采用FPGA+ MSP430双核架构，nRF905无线模块传输距离150m，整机功耗290mW。采集的氧化钐粉末吸光度光谱与台式光谱仪完全吻合。替代设计提出反射式方案：微透镜阵列+反光杯+柱面镜，尺寸Φ24.6×34.3mm。ZEMAX优化后光斑能量集中度85%在中心像元，TracePro追光分析显示杂散光占比低于3%。户外实验验证节点可采集50m外树叶反射光谱，信噪比优于200:1。
import numpy as np from scipy.signal import savgol_filter from sklearn.cross_decomposition import PLSRegression from sklearn.svm import SVC from sklearn.model_selection import train_test_split import pywt def snv_transform(spectrum): return (spectrum - np.mean(spectrum)) / np.std(spectrum) def mw_bpann_feature_selection(X, y, window_width=30, step=15): n_wavelengths = X.shape[1] accuracies = [] for start in range(0, n_wavelengths - window_width, step): window = X[:, start:start+window_width] # simple dummy classifier for demo acc = np.random.rand() * 0.2 + 0.7 accuracies.append((start+window_width//2, acc)) best_center = max(accuracies, key=lambda x: x[1])[0] return best_center def mutagenic_genetic_algorithm(X, y, n_generations=5, mutation_rate=0.15): n_features = X.shape[1] pop_size = 30 population = np.random.rand(pop_size, n_features) > 0.9 for gen in range(n_generations): fitness = [] for indiv in population: selected = X[:, indiv] if selected.shape[1] == 0: fitness.append(0) continue # dummy fitness based on SVM accuracy acc = 0.75 + 0.15 * (np.sum(indiv)/n_features) fitness.append(acc) # selection and mutation new_pop = [] for _ in range(pop_size): parent = population[np.argmax(fitness)] child = parent.copy() if np.random.rand() < mutation_rate: idx = np.random.randint(n_features) child[idx] = not child[idx] new_pop.append(child) population = np.array(new_pop) best = population[np.argmax(fitness)] return np.where(best)[0] def design_optical_system(distance=50, fov=10.0, detector_width=0.5): # thin lens equation focal_length = (detector_width/2) / np.tan(np.radians(fov/2)) magnification = focal_length / distance return focal_length, magnification def simulate_throughput(aperture=10, transmittance=0.85): return aperture * transmittance if __name__ == '__main__': # Simulate NIR spectra np.random.seed(42) n_samples = 200 n_wav = 156 X = np.random.randn(n_samples, n_wav) * 0.1 + 0.5 y = np.random.randint(0,2, n_samples) # binary class X_snv = snv_transform(X[0]) print(f'SNV transform range: [{X_snv.min():.2f}, {X_snv.max():.2f}]') best_wl = mw_bpann_feature_selection(X, y) print(f'MW-BPANN selected wavelength index: {best_wl}') selected_idx = mutagenic_genetic_algorithm(X, y, n_generations=3) print(f'Mutagenic GA selected {len(selected_idx)} features') pls = PLSRegression(n_components=8) pls.fit(X[:150], y[:150]) y_pred = pls.predict(X[150:]) print(f'PLS-DA prediction accuracy: {np.mean((y_pred>0.5).flatten() == y[150:]):.3f}') focal, mag = design_optical_system(distance=50, fov=12) print(f'Optical system: focal length={focal:.1f}mm, magnification={mag:.5f}') throughput = simulate_throughput() print(f'Optical throughput: {throughput:.2f} mm²')
