Scheme 1. Schematic illustration of the construction and application of the CWA-UMAP-LDA-assisted RFLSA for St and its structural analogues in water.#

Fig.1 (a, b) TEM images of the FB220-doped Tb/Eu-MOFs. (c) SEM image of the FB220-doped Tb/Eu-MOFs. The inset was an OM image of the MOFs dispersed in aqueous solution. (d and e) SEM and the corresponding elemental mapping images of FB220-doped Tb/Eu-MOFs. (f) FTIR spectra of Tb/Eu-MOFs, FB220, and FB220-doped Tb/Eu-MOFs. (g) XRD patterns of Tb/Eu-MOFs and FB220-doped Tb/Eu-MOFs. (h) UV-Vis absorption (curve 1), emission (curve 2, excitation wavelength at 265 nm), and excitation spectra (curve 3, emission wavelength at 542 nm) of the FB220-doped Tb/Eu-MOFs.#

在响应机制探究中，如图2所示，探针单元与苯乙烯及肉桂腈、白藜芦醇衍生物等七种结构类似物发生交叉响应，产生了独特的差异化荧光指纹图谱。探针1的响应主要受内滤效应和分析物本征发光特性的调控，而探针2则引入了高锰酸钾对分析物的氧化反应，通过荧光信号的猝灭与恢复实现开关式调控。由于不同探针单元对特定目标物的分辨能力存在差异，直接拼接数据易削弱强响应通道的贡献。为此，研究团队引入了通道权重分析对探针数据进行缩放处理，将探针1与探针2的权重比例优化为1<9>。

Fig.2 (a, b) The FL emission spectra of P1 and P2 with 5 mg/L of eight analytes, respectively. (c) Bar plots from the proposed RFLSA for each target. (d) PCA, (e) LDA, (f) ISOMAP, (g) MDS, (h) UMAP, and (i) UMAP-LDA plot of eight targets identification.#

为了高效提取高维光谱数据的判别特征，研究对比了多种模式识别算法。如图2与图3所示，传统的主成分分析与线性判别分析在复杂体系中存在重叠或类间距离较近的问题，而采用连续UMAP-LDA工作流（即均匀流形近似与投影联合线性判别分析）后，能够完美保留高维数据的局部拓扑结构与全局分布特征。优化后的通道权重分析联合连续工作流展现出卓越的类间分离能力，热图、层次聚类分析图与混淆矩阵验证了该模型在分类预测中的100%准确率。

Fig.3 UMAP-LDA plots of eight analytes (5 mg/L) with different weighting ratios, including (a) 2<8>, (b) 8<2>, (c) 3<7>, (d) 7<3>, (e) 4<6>, (f) 6<4>, (g) 9<1>, and (h) 1<9>. (i) heatmap, (j) HCA plot, (k) radar plot, and (l) confusion matrix for the RFLSA with eight analytes via CWA-UMAP-LDA.#

在定量检测性能评估中，如图4所示，探针2在425 nm与542 nm处的比率荧光信号强度与苯乙烯及七种类似物浓度呈现出优异的线性关系。苯乙烯的最低检出限达到0.001 mg/L，显著低于世界卫生组织的饮用水标准。同时，该传感器阵列对不同浓度梯度的类似物均实现了清晰的簇类划分。针对多组分共存的复杂水样，如图5所示，传感器阵列成功对二元至八元混合物（如苯乙烯/4-氯肉桂酸混合体系）在不同配比下进行了精准区分，展示出极强的抗干扰能力与模式识别鲁棒性。

Fig.4 (a) Waterfall plot of FL response for P2 in the presence of St with different concentrations. (b) The relationship of the I425/I542 values to St concentrations. The inset displayed the standard curve between I425/I542 and St concentration C (0.001–2 mg/L). CWA-UMAP-LDA plots (c) and HCA (d) based on FL response patterns towards varying concentrations of St. (e-h) CWA-UMAP-LDA plots of St analogues at concentrations of (e) 0.01, (f) 0.1, (g) 1, and (h) 2 mg/L.#

Fig.5  LDA plots of the CWA-UMAP-LDA-based RFLSA for the mixtures at a total concentration of 4 mg/L, including (a) binary, (c) ternary, (e) quaternary, and (g) quinary-to-octonary mixtures. (b), (d), (f), and (h) were the corresponding HCA plots.#

实际水样分析验证了该技术的实用价值。如图6与表1所示，在白河水与自来水加标实验中，八种分析物的回收率达到90.30%–109.65%，且所有盲样均被正确归类。在针对真实塑料厂废水的检测中，气相色谱-质谱法测得苯乙烯浓度为0.66 mg/L，经20倍稀释至0.033 mg/L后，传感器阵列与数据处理模型能够精准辨识并归级该废水中分析物的响应特征，证实了该平台在真实环境水体监控中的可靠性。

Fig.6 CWA-UMAP-LDA (a, c, e, and g) and HCA plots (b, d, f, and h) derived from FL responses of the RFLSA toward eight analytes in Baihe River and tap water, respectively. (i) CWA-UMAP-LDA and (j) HCA plots based on FL response patterns for factory wastewater containing St (0.033 mg/L). The USs were distinguished using different colors and shapes.#