
3D fluorescence spectroscopy is commonly used for water pollutant identification and analysis, but traditional systems for acquiring and processing 3D fluorescence spectra are not suitable for portable instruments with limited resources due to the bulky optical scanning device, large data volume and computational complexity. In this paper, a portable water pollutant detection system based on 3D fluorescence spectra with a sparse reconstruction-encoding strategy is proposed. In the system, a multi-wavelength LED-filter synergy module was designed to capture sparse 3D fluorescence spectra. Additionally, a super-resolution reconstruction algorithm based on dictionary learning is integrated, enhancing the dimensionality and information of sparse data. To make the algorithm lightweight for direct implementation on the instrument, centroid-guided feature points are extracted in layers, with an efficient and lightweight joint encoding method introduced to consider both positional and intensity correlations between feature points. Then, a Histogram Intersection Kernel Support Vector Machine is implemented on an embedded platform for identification of water pollutants. Validation experiments on six common water pollutants show that the proposed method outperforms conventional approaches, offering an innovative and practical solution for portable on-site water pollutant detection.
Multipath interference is a key factor affecting the reception performance of regional navigation signals. The Multipath Estimating Delay Locked Loop (MEDLL) can simultaneously estimate the relative amplitude, chip delay, and phase offset of direct and multipath signals, and has been widely used in anti-multipath tracking loops. The actual working environment inevitably contains noise, which can distort multipath correlation functions, leading to increased estimation errors of MEDLL parameters. To address this issue, an improved MEDLL algorithm combining Complementary Ensemble Empirical Mode Decomposition (CEEMD) and Sample Entropy (SpEn) is proposed, namely the CEEMDSpEn MEDLL algorithm. This algorithm utilizes SpEn to screen the effective signal components after CEEMD decomposition, reconstructs the multipath correlation function, and effectively suppresses the distortion of the correlation function curve caused by noise, thereby improving the accuracy of multipath parameter estimation under noise interference. The experimental results show that the multipath parameter estimation error of the proposed algorithm is smaller than that of the traditional MEDLL method at signal-to-noise ratios of 25 dB to 5 dB.
Artificial intelligence (AI) systems are increasingly used in medical imaging tasks such as lesion detection, tumor classification, and disease grading [1]. While these systems often achieve high predictive performance, their integration into clinical workflows is still limited by the lack of measurable and operational definitions of trust. From an Instrumentation and Measurement (I&M) perspective, modern AI systems can be viewed as complex, non-linear measurement instruments that transform raw medical data, such as magnetic resonance images (MRI) or X-ray computed tomography (CT) images, into diagnostic outputs. In this analogy, a deep learning model like the convolutional neural network (CNN) functions as a multi-stage measurement system, where each layer performs successive transformations of the input signal into increasingly abstract representations, with the final output corresponding to a measurement decision. Unlike traditional instruments, however, the internal states of AI systems are not directly observable and lack standardized calibration and traceability procedures. Thus, their reliability is typically assessed only through external performance metrics such as accuracy or sensitivity. This opacity motivates a measurement-oriented interpretation in which components of AI systems can be mapped to classical I&M elements: input data correspond to measured signals, network layers to signal conditioning stages, learned features to internal instrument states, and training dynamics to a calibration process. Within this framework, trust is considered in relation to measurable properties of the stability and consistency of internal representations rather than as a purely qualitative assessment.
Motivated by the RUL prediction requirements of rolling bearings in coke pushing systems, this paper addresses the difficulty of effectively extracting degradation features of rolling bearings under different load and rotational-speed conditions, as well as the limited generalization capability of traditional data-driven models, and proposes a physical prior enhanced remaining useful life prediction method. Specifically, for the evolution process of local defects, two independent matching objectives are formulated in the time and frequency domains by combining the bearing dynamic model with measured vibration signals, using time domain composite statistical indicators and frequency-domain high-order harmonic amplitude features, respectively, thereby quantifying defect-related physical prior information. Meanwhile, the quantified multi-view physical priors are fused with optimized data-driven degradation features, and a multi-scale dynamic network is designed. Through multi-scale convolution and bidirectional gated recurrent unit, the network extracts spatiotemporal degradation patterns, captures local spatial features and global temporal dependencies, and realizes RUL prediction for rolling bearings. Finally, experiments conducted on two bearing datasets demonstrate that introducing equivalent-defect-size-related physical priors can improve the prediction performance of the proposed RUL prediction model under the selected experimental conditions, while incorporating defect-related physical information into the prediction framework, providing methodological support for its further application to rolling bearings in coke pushing systems.
The reliability of hyperspectral analysis does not begin with the algorithm, but with the sensor. Hyperspectral Imaging (HSI) has evolved from a specialized tool into a widely used measurement technology, capable of extracting optical signatures in many different settings, ranging from industry and agriculture to crime scenes. However, the rapid adoption of these systems masks a fundamental methodological issue: HSI sensors are frequently used by end-users as infallible black boxes. The state of the art focuses on downstream algorithmic processing and sophisticated predictive models and neglects the crucial role of hardware instrumentation. On a representative testbed, we show that routine, undocumented acquisition choices alone account for an expanded uncertainty of 6.41% (K = 2) on the extracted spectral quantity, a worst-case figure obtained with no spatial correction applied and one that is already present in the raw data before any predictive algorithm is used.