A trace volatile non-contact detection system and method based on multispectral array sensing

By combining multispectral array sensing technology and dynamic environmental compensation model, high-sensitivity and fast-response trace volatile detection is achieved in complex environments, solving the problems of insufficient accuracy and response speed in existing technologies, and is suitable for high-throughput scenarios such as airports.

CN122171466APending Publication Date: 2026-06-09YUNNAN POLICE COLLEGE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610285520.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-10
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

Existing trace substance detection technologies are insufficient in terms of accuracy, sensitivity, anti-interference and response speed, making it difficult to meet the high-throughput screening needs of complex scenarios such as airports and train stations.

Method used

Employing multispectral array sensing technology, combined with an inert material detection chamber, an environmental sensing unit, and a dynamic environmental compensation model, it achieves rapid and accurate detection of trace volatiles through non-contact sampling and intelligent signal processing.

Benefits of technology

It achieves high sensitivity, anti-interference and fast response in trace substance detection, with a false alarm rate of less than 2%, and is suitable for high-throughput screening needs in complex environments. The hardware is portable and easy to operate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122171466A_ABST
    Figure CN122171466A_ABST
Patent Text Reader

Abstract

The application discloses a trace volatile non-contact detection system and method based on a multispectral array sensor, and the trace volatile non-contact detection system based on the multispectral array sensor comprises a gas collection unit, which is suitable for introducing an air sample of a to-be-detected region into the system in a non-contact mode; and a detection cavity, which is in communication with the gas collection unit, and the inner wall of the detection cavity is made of an inert material. The application realizes ultrahigh-sensitivity detection, can accurately capture trace target substance volatiles, and a multispectral sensing array is subjected to specific band optimization for characteristic absorption peaks of the target substance, thereby abandoning a redundant design of continuous full-spectrum scanning, greatly improving the signal-to-noise ratio for specific targets, and cooperating with a multi-dimensional dynamic environment compensation mechanism to accurately correct signal baseline drift and nonlinearly adjust sensitivity, so that the detection level is far superior to that of a traditional electronic nose technology, and trace target substance volatiles in air can be effectively captured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of public safety technology, specifically to a non-contact detection system and method for trace volatiles based on multispectral array sensing. Background Technology

[0002] Existing trace substance detection technologies are mainly divided into three categories, and their specific shortcomings are as follows: 1. Electronic nose detection technology: This technology simulates the biological olfactory mechanism, using a non-specific gas sensor array for odor "pattern recognition." While it enables non-contact gas sampling, it does not inherently involve precise chemical component identification. Because the sensor array lacks specific adaptation to the characteristic wavelengths of the target substance, it is highly susceptible to interference from perfumes, alcohol, food volatiles, and exhaust fumes in complex chemical environments such as airports and train stations, leading to signal confusion. More importantly, this technology is extremely sensitive to changes in environmental temperature and humidity; temperature fluctuations or increased humidity can cause sensor baseline drift, significantly reducing the stability of detection data, resulting in a high false alarm rate. It cannot accurately distinguish trace volatiles from environmental interference signals, failing to meet the sensitivity requirements for trace-level detection.

[0003] 2. Laboratory analytical techniques (such as gas chromatography-mass spectrometry, GC-MS): As the "gold standard" for confirming substance composition, this technique achieves extremely high detection accuracy through complex chemical separation and mass spectrometry analysis, but it has significant limitations in application. Its core drawbacks lie in the large size of the equipment, high purchase and maintenance costs, and cumbersome operation procedures requiring specialized technicians for sample processing and data analysis. More importantly, this technology relies on the sample delivery process, which can take hours or even days from sampling to obtaining test results, completely failing to meet the real-time screening needs of scenarios such as airports and border inspections. It can only serve as a post-event confirmation method and cannot provide early warning.

[0004] 3. Portable on-site detection technologies (such as ion mobility spectrometry (IMS) and Raman spectroscopy): These technologies are currently the mainstream on-site detection solutions, but they have fundamental limitations. On the one hand, their detection methods rely on contact sampling, requiring the acquisition of physical samples by wiping the surface of suspicious items or human skin. They are completely unable to detect volatile substances that leave no trace, thus failing to achieve non-contact early warning. On the other hand, the detection principle of IMS technology is easily affected by other volatile organic compounds in the environment, resulting in a high false alarm rate. Furthermore, the sampling, sample pretreatment, and analysis processes are relatively cumbersome, requiring professionally trained personnel to operate. The detection speed is limited by process efficiency, making it difficult to adapt to the rapid screening needs of high-throughput pedestrian and logistics scenarios.

[0005] In addition, all existing target substance detection technologies have certain shortcomings in some aspects. Electronic nose technology can achieve non-contact detection, but it has the defect of insufficient sensitivity. Laboratory analysis methods have high sensitivity, but require sampling and testing, which takes a long time and cannot achieve real-time on-site detection. On-site portable detection technologies usually require targeted contact detection.

[0006] In view of this, the present invention provides a comprehensive solution for the detection of trace chemical substances that combines multispectral chip sensing technology, advanced machine learning algorithms and dynamic environmental compensation technology. It has the characteristics of fast response speed, insensitivity to environmental interference and good sensitivity, and provides a reliable tool for the prevention and control of target substances in the field of public safety. Summary of the Invention

[0007] The purpose of this invention is to provide a non-contact detection system and method for trace volatiles based on multispectral array sensing, so as to solve the problems existing in the prior art mentioned in the background.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a non-contact detection system for trace volatiles based on multispectral array sensing, comprising: A gas sampling unit is adapted to introduce an air sample from the area to be tested into the system in a non-contact manner; The detection chamber is connected to the gas collection unit, and the inner wall of the detection chamber is made of an inert material, which is suitable for providing a stable optical detection environment. A multispectral sensing array is disposed on the side wall of the detection cavity, and the multispectral sensing array includes several sets of photoelectric sensing units optimized for the characteristic wavelength bands of the target substance. An environmental sensing unit is disposed within the detection cavity, and the environmental sensing unit is adapted to collect temperature and humidity parameters of the detection environment in real time. A multi-band light source assembly is disposed in the detection cavity on the side away from the multispectral sensing array. The multi-band light source assembly is adapted to emit detection light covering the ultraviolet, visible and near-infrared bands. The processing and control unit is electrically connected to the gas acquisition unit, the multispectral sensing array, the environmental sensing unit, and the multi-band light source assembly, respectively. The processing and control unit is equipped with an intelligent signal processing module, which is configured to perform the following operations: Acquire the original time-series signal output by the multispectral sensing array; Using the temperature and humidity parameters, the original signal is baseline drift corrected through a preset dynamic environmental compensation model; A density-based clustering algorithm is used to clean outliers in the corrected signal, removing noisy data points with a density below a preset threshold. The cleaned feature data is input into a pre-trained ensemble learning model for inference, and the model outputs the category and concentration information of the target substance.

[0009] Preferably, the inner wall material of the detection chamber is polytetrafluoroethylene (PTFE) or perfluoroalkoxy resin (PFA), and the air inlet and outlet of the detection chamber are streamlined to reduce gas adsorption and dead volume.

[0010] Preferably, the multispectral sensing array comprises 32 independent photoelectric sensing units, the response bands of which are discontinuously distributed and specifically cover specific characteristic peak bands in the 200nm-400nm ultraviolet region, the 400nm-700nm visible light region, and the 700nm-1700nm near-infrared region.

[0011] Preferably, the multi-band light source component is driven by pulse modulation, and the processing control unit is configured to acquire signals only during the pulse period when the light source is turned on, so as to eliminate background light interference.

[0012] A non-contact detection method for trace volatiles based on multispectral array sensing, applied to any of the aforementioned detection systems, includes the following steps: S1) Data Acquisition The gas acquisition unit is activated to simultaneously acquire the multi-channel response signals of the multispectral sensor array and the temperature and humidity parameters of the environmental sensing unit. S2) Dynamic Environmental Compensation Based on the temperature and humidity parameters, zero-baseline correction and sensitivity nonlinearity compensation are performed on the multi-channel response signal. S3) Adaptive Cleaning The compensated multi-channel response signal is mapped to a high-dimensional feature space, and discrete noise points are identified and removed by density clustering algorithm. S4) Feature Filtering Calculate the feature importance of each spectral channel for the identification of the target substance, and select key channel combinations as input features; S5) Intelligent Recognition The filtered input features are fed into the ensemble learning model, which integrates the prediction results of multiple classifiers and outputs the detection results and classification confidence.

[0013] Preferably, in step S2, dynamic environmental compensation includes the following sub-steps: S21) Baseline Drift Correction Based on the real-time collected temperature and humidity parameters, the corresponding zero-point drift vector is retrieved from the preset database, and the zero-point drift vector is subtracted from the multi-channel response signal. S22) Sensitivity Correction The absorption suppression coefficient of water molecules on a specific wavelength band is calculated based on the real-time humidity parameters, and the signal strength of a specific channel is compensated and amplified using a nonlinear gain function. The preset database contains zero-point drift vectors corresponding to multiple temperature and humidity combinations within the range of -10 to 50°C and 10% to 95% humidity. The nonlinear gain function is suitable for counteracting the suppressive effect of ambient humidity on detection sensitivity.

[0014] Preferably, in step S3, the adaptive cleaning uses the DBSCAN density clustering algorithm, which includes the following sub-steps: S31) Set the neighborhood radius (eps) and minimum number of samples (MinPts); S32) Traverse all data points in the high-dimensional feature space and calculate the number of neighboring points of each data point within the neighborhood radius; S33) If the number of neighboring points of a data point is greater than or equal to the minimum number of samples, then mark it as a core point and cluster it with its density-reachable neighboring points into one class; S34) If a data point is neither a core point nor in the neighborhood of any core point, then mark it as an outlier noise point and remove it from the signal sequence or perform interpolation smoothing. The neighborhood radius (eps) ranges from 0.1 to 0.5, the minimum number of samples (MinPts) ranges from 50 to 200, and the algorithm is suitable for removing interference data caused by occasional circuit noise or airflow jitter.

[0015] Preferably, in step S4, feature selection includes the following sub-steps: S41) During the model training phase, the Gradient Boosting Classifier is used to fit the training data; S42) Calculate the average information gain for each spectral channel during decision tree node splitting. S43) Channels with information gain below a preset threshold are defined as redundant channels and removed, while the top K channels in terms of information gain are retained as key feature channels. The value of K ranges from 5 to 15, and the cumulative contribution of information gain of the key feature channels is not less than 85%, which is suitable for reducing the model's computational dimensionality and improving inference efficiency.

[0016] Preferably, in step S5, the intelligent recognition adopts a Stacking ensemble learning model, which includes the following sub-steps: S51) Construct the Stacking ensemble learning model, in which at least two gradient boosting decision tree algorithm models selected from LightGBM, XGBoost or CatBoost are deployed in parallel for the first layer base learner, and the second layer meta learner adopts logistic regression or a linear model. S52) Input the filtered key feature channel data into the first layer base learner, and each base learner outputs a prediction probability vector; S53) The second-layer meta-learner receives the predicted probability vectors from each base learner and outputs the detection result and classification confidence of the target substance through weighted fusion; The first layer base learners are LightGBM, XGBoost, and CatBoost algorithm models, respectively. The training data of each model are divided using cross-validation to improve the stability and accuracy of the detection results.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1) This application achieves ultra-high sensitivity detection, which can accurately capture trace amounts of target substance volatiles. The multispectral sensing array is specifically optimized for the characteristic absorption peaks of the target substance, abandoning the redundant design of continuous full-spectrum scanning, and significantly improving the signal-to-noise ratio for specific targets. Combined with a multi-dimensional dynamic environmental compensation mechanism to accurately correct signal baseline drift and non-linearly adjust sensitivity, the detection limit reaches 0.01 micrograms per cubic meter, which is far superior to the detection level of traditional electronic nose technology. Thus, it can effectively capture trace amounts of target substance volatiles in the air and avoid the risk of missed detection from the source. 2) This application has strong anti-interference capabilities, which can ensure the detection accuracy in complex environments. The detection chamber is made of inert material and combined with the streamlined air inlet and outlet design to reduce gas adsorption and dead volume, and avoid the memory effect caused by residual volatiles. The dynamic environmental compensation mechanism is based on a zero-point drift database covering a wide temperature and humidity range. It generates a real-time zero-point baseline vector through interpolation algorithm to accurately offset the influence of temperature and humidity fluctuations on the detection signal. The adaptive cleaning algorithm can identify and remove outlier noise points caused by circuit jitter, airflow changes, etc. The combined effect of multiple anti-interference designs makes the system false alarm rate less than 2%, effectively solving the problem that the existing technology is easily interfered with and the detection results are unstable in complex chemical background environments, ensuring the detection accuracy in complex scenarios such as airports, stations, and entertainment venues. 3) This application achieves a perfect combination of rapid response and non-invasive detection, adapting to diverse high-throughput scenario requirements. The gas collection unit introduces air samples through active or passive non-contact methods, eliminating the need to obtain physical samples. This avoids interference with normal pedestrian and material flow caused by contact sampling, while also detecting volatile substances without leaving any traces. It can achieve covert detection without affecting the order of the site, adapting to the safety screening needs of various high-risk areas. At the same time, the feature screening algorithm evaluates the information gain of each channel, eliminates redundant bands, and significantly reduces the computational dimension. Combined with the efficient reasoning capability of the integrated learning model, the entire detection process from sampling to result output does not exceed 5 seconds, which is much faster than traditional detection technologies. It can meet the real-time screening needs of high-throughput scenarios such as airports and train stations, balancing detection efficiency and ease of use. 4) The hardware of this application is portable, reliable, and easy to operate, significantly reducing the threshold for technology implementation and promotion. The optimized hardware structure controls the detection chamber volume to within 20 cubic centimeters, supports battery power, and can be deployed in various scenarios without complicated installation and debugging. It solves the problems of traditional laboratory equipment being bulky, poorly portable, and dependent on fixed power sources. The system integrates automatic cleaning, remote monitoring, and alarm output functions. Operators can use it without professional training, which reduces the manpower cost of security screening and can quickly adapt to the usage needs of different application scenarios, improving the feasibility of technology promotion and application, and providing an efficient and practical tool for trace substance screening in the field of public safety. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent algorithm processing in this application; Figure 2 This is a cross-sectional view of the system device structure of this application; Figure 3 This is a flowchart of the testing method of this application; Figure 4 This is a step-by-step flowchart of the dynamic environmental compensation process in this application; Figure 5 This is a step-by-step flowchart of the adaptive cleaning process in this application. Figure 6 Step-by-step flowchart for feature selection in this application Figure 7 This is a step-by-step flowchart of the intelligent recognition process in this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of the invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front end," "rear end," "both ends," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0021] In the description of the invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] In the description of the invention, it should be noted that the execution order of the steps is not limited by the sequence number. The possible changes in the order of some steps, the synchronous execution of steps, and the split execution of steps are all within the scope of protection of this application.

[0023] Please see Figure 1-7 This invention provides a technical solution: a non-contact detection system for trace volatiles based on multispectral array sensing, comprising: The gas sampling unit is suitable for introducing air samples from the area to be tested into the system in a non-contact manner; The detection chamber is connected to the gas acquisition unit. The inner wall of the detection chamber is made of inert material, which is suitable for providing a stable optical detection environment. A multispectral sensing array is set on the side wall of the detection cavity. The multispectral sensing array contains several sets of photoelectric sensing units optimized for the characteristic wavelength bands of the target material. An environmental sensing unit is installed inside the detection chamber. The environmental sensing unit is suitable for real-time acquisition of temperature and humidity parameters of the detection environment. The multi-band light source assembly is located on the side of the detection cavity away from the multispectral sensing array. The multi-band light source assembly is suitable for emitting detection light covering the ultraviolet, visible and near-infrared bands. The processing and control unit is electrically connected to the gas acquisition unit, the multispectral sensor array, the environmental sensing unit, and the multi-band light source assembly. The processing and control unit is equipped with an intelligent signal processing module, which is configured to perform the following operations: Acquire the raw time-series signal output from the multispectral sensor array; Using temperature and humidity parameters, the baseline drift of the original signal is corrected through a preset dynamic environmental compensation model; A density-based clustering algorithm is used to clean outliers in the corrected signal, removing noisy data points with a density below a preset threshold. The cleaned feature data is input into a pre-trained ensemble learning model for inference, and the model outputs the category and concentration information of the target substance.

[0024] Specifically, the non-contact design of the gas collection unit allows for air sample collection without wiping or touching suspicious objects, whereas existing technologies rely on physical sample acquisition, which has limitations. This design avoids interference with normal pedestrian and material flow while detecting volatile substances without leaving any trace, making it suitable for covert screening needs in various high-risk scenarios such as airports and entertainment venues. The detection chamber is made of inert materials, providing a stable and interference-free environment for optical detection and reducing the impact of external factors on the signal from the source. The multispectral sensing array is optimized for the characteristic wavelengths of the target substance, eliminating the defects of non-specific sensing and laying the foundation for accurate identification. The environmental sensing unit collects temperature and humidity parameters in real time, providing data support for subsequent dynamic compensation and solving the problem of the lack of environmental adaptation mechanism in existing technologies. The processing control unit is electrically connected to each component and integrates an intelligent signal processing module. Through baseline drift correction, density clustering cleaning, and hierarchical processing of integrated learning inference, it achieves accurate signal optimization and efficient identification, effectively making up for the core shortcomings of traditional detection technologies, such as low sensitivity, weak anti-interference, and slow response.

[0025] The inner wall material of the detection chamber is polytetrafluoroethylene (PTFE) or perfluoroalkoxy resin (PFA), and the air inlet and outlet of the detection chamber are streamlined to reduce gas adsorption and dead volume. Specifically, this application solves the problem of existing detection chambers easily adsorbing volatile organic compounds and causing detection errors due to gas retention. The PTFE or PFA used in the detection chamber are typical inert materials with extremely low surface energy, which can effectively avoid the "memory effect" formed by the adsorption of volatile organic compounds on the chamber wall, ensuring that each detection is not affected by the residue of the previous sample, and significantly improving the stability and accuracy of the detection results. The streamlined design of the air inlet and outlet can reduce the residence time and dead volume of gas in the chamber, accelerate the gas exchange rate, improve the detection response efficiency, and avoid concentration distortion caused by gas accumulation, which is especially suitable for the continuous detection needs in rapid screening scenarios.

[0026] The multispectral sensing array comprises 32 independent photoelectric sensing units. The response bands of these units are discontinuously distributed and specifically cover specific characteristic peak bands in the 200nm-400nm ultraviolet, 400nm-700nm visible, and 700nm-1700nm near-infrared regions. Specifically, the 32 independent photoelectric sensing units achieve comprehensive coverage of specific characteristic peak bands in the ultraviolet, visible, and near-infrared regions, enabling precise capture of characteristic absorption signals from volatile components of the target substance, ensuring effective identification of trace substances. The discontinuously distributed response band design eliminates redundant data from irrelevant bands in full-spectrum scanning, significantly reducing data processing load while improving the signal-to-noise ratio for the target substance, thus allowing for rapid inference in subsequent algorithms. The specifically covered characteristic peak bands enable the sensing array to accurately distinguish the signal differences between the target substance and environmental interferences, further improving detection accuracy.

[0027] The multi-band light source component is driven by pulse modulation, and the processing control unit is configured to acquire signals only during the pulse period when the light source is on, in order to eliminate background light interference. Specifically, the multi-band light source component is driven by pulse modulation, which can concentrate the emission of detection light within a specific pulse period, and the processing control unit only acquires signals within that pulse period. This is equivalent to setting a "precise window" for signal acquisition, effectively avoiding continuous interference from background light such as natural light and artificial light in the environment, and significantly improving the purity of the acquired signal. This design eliminates the need for additional light-shielding equipment, simplifying the hardware structure while ensuring detection stability in complex lighting scenarios such as airports and train stations.

[0028] Specifically, in implementation, the gas collection unit can be a gas collector, which includes a gas inlet and a gas pump. By installing the gas collector in high-risk areas such as train stations, airports, bars, and entertainment venues, non-invasive gas sample collection can be achieved. Gas collectors can be divided into active and passive collection. Active collection requires a gas pump to assist in drawing in the sample gas and delivering it to the sampling chamber; passive collection does not require a gas pump and uses a static installation with the gas inlet, allowing the sample gas to flow naturally into the inlet.

[0029] Specifically, in implementation, acquiring multi-band spectral data involves a detection chamber, a multi-band light source assembly, and a multi-spectral sensing array. The gas acquisition unit delivers the collected sample gas into the detection chamber. The inner wall of the detection chamber is made of an inert material (such as polytetrafluoroethylene) to prevent VOCs adsorption from affecting the detection results. After the sample gas receives illumination from the multi-band light source assembly within the detection chamber, a discontinuous multi-spectral array sensing assembly optimized for specific functional groups is installed on the other side of the detection chamber. This assembly does not uniformly cover the entire spectrum; instead, based on the characteristic absorption peak positions of the target substance (such as ester groups, benzene ring structures, and amine groups), narrowband photosensitive units specifically selected include key wavelengths such as 220nm (ultraviolet characteristic region), 450nm (visible light sensitive region), and 950nm and 1050nm (near-infrared frequency doubling region). Compared to continuous full-spectrum scanning, this "discrete array" design significantly improves the signal-to-noise ratio for specific targets while reducing data processing requirements. It enables real-time acquisition of spectral signals of volatile organic compounds (VOCs) in the air, achieving high-sensitivity (detection limit down to 0.01 μg / m³) capture of trace substances.

[0030] Specifically, in implementation, the intelligent data processing and recognition algorithm adopts a hierarchical data processing strategy. The first level is density-based adaptive data cleaning: addressing the issue of outlier noise generated by sensors in complex environments due to transient interference (such as wind pressure changes and circuit vibrations), DBSCAN (density-based clustering algorithm) is introduced. By setting a density threshold and neighborhood radius, the density distribution of multidimensional spectral data in the feature space is calculated, automatically identifying and removing isolated noise points that do not conform to the characteristics of continuous chemical reactions, outputting a clean signal with a high signal-to-noise ratio. The second level is feature engineering and dimensionality reduction: using gradient boosting tree (such as GBDT) algorithms to evaluate the feature importance of 32-channel spectral data, automatically selecting the key band combinations that contribute the most to the target substance, removing redundant bands, and reducing the computational dimensionality. The third level is ensemble learning: abandoning single models, an ensemble learning model based on a stacking strategy is constructed. The prediction results of multiple heterogeneous classifiers such as LightGBM, XGBoost and CatBoost are fused by a meta-learner and output qualitative recognition results and confidence scores in milliseconds.

[0031] Specifically, in implementation, the multi-dimensional dynamic environmental compensation mechanism goes beyond simple numerical deduction; it constructs a "temperature and humidity-spectral baseline nonlinear correction model." Furthermore, the system pre-loads a zero-point drift database for the sensor array under different temperature and humidity combinations (temperature -10~50℃, humidity 10%~95%). During detection, based on real-time acquired temperature and humidity data, an interpolation algorithm dynamically generates the current "zero-point baseline vector," performing differential correction on the original spectral data. Finally, considering the changes in gas molecule adsorption dynamics under high temperature and humidity conditions, the algorithm dynamically adjusts the gain coefficient of specific characteristic bands using a nonlinear function based on the current absolute humidity value to counteract the suppressive effect of the environment on sensitivity, ensuring high-sensitivity detection even in extreme environments.

[0032] Example 1: Hardware Construction of the Detection System This embodiment provides a non-contact detection system for trace volatiles based on multispectral array sensing. The various components of its hardware structure work together to achieve non-contact trace detection functionality, and the specific construction is as follows: 1. Detection Chamber Design: The streamlined detection chamber is made of polytetrafluoroethylene (PTFE) material, with a chamber volume strictly controlled within 20 cm³ to minimize dead volume. PTFE material has extremely low surface energy, effectively preventing the "memory effect" caused by the adsorption of volatile organic compounds (VOCs) on the chamber walls, ensuring the accuracy of the detection results. An air inlet and outlet are respectively located on both sides of the chamber. A miniature brushless air pump is connected to the air inlet, precisely controlling the airflow velocity at 500 mL / min to achieve stable non-contact gas sampling.

[0033] 2. Multispectral Sensing Array Configuration: A 32-channel (A1-A32) photoelectric sensor array (i.e., multispectral sensing array) is integrated and installed on one inner wall of the detection cavity. This array adopts a discrete optimization design, not covering the continuous spectrum, but precisely adapted to the characteristic absorption peaks of the target substance: 5 channels are configured in the ultraviolet region (200-400nm) to match the strong absorption characteristics of benzene ring structures; 12 channels are configured in the visible light region (400-700nm) to adapt to colorimetric reactions and specific functional group recognition; and 15 channels are configured in the near-infrared region (700-1700nm) to target the overtone absorption characteristics of CH and NH bonds. Simultaneously, a high-precision temperature and humidity sensor (SHT series) is embedded in the inner wall of the cavity to collect the cavity's temperature and humidity parameters in real time, providing data support for subsequent dynamic environmental compensation.

[0034] 3. Light Source Component Selection: A multi-band LED light source board (i.e., a multi-band light source component) is installed on the opposite side of the multispectral sensing array. The light source is driven by pulse modulation, with the modulation frequency set to 10Hz. This design allows the processing control unit to acquire signals only during the pulse cycle when the light source is on, effectively eliminating interference from ambient background light on the detection signal and improving the purity of the original signal.

[0035] According to another aspect of this application, a non-contact detection method for trace volatiles based on multispectral array sensing is also provided, applied to a non-contact detection system for trace volatiles based on multispectral array sensing, comprising the following steps: S1) Data Acquisition The gas acquisition unit is activated to simultaneously acquire the multi-channel response signals of the multispectral sensor array and the temperature and humidity parameters of the environmental sensing unit. S2) Dynamic Environmental Compensation Based on temperature and humidity parameters, zero-baseline correction and sensitivity nonlinearity compensation are performed on the multi-channel response signal; S3) Adaptive Cleaning The compensated multi-channel response signal is mapped to a high-dimensional feature space, and discrete noise points are identified and removed by density clustering algorithm. S4) Feature Filtering Calculate the feature importance of each spectral channel for the identification of the target substance, and select key channel combinations as input features; S5) Intelligent Recognition The filtered input features are fed into the ensemble learning model, which combines the prediction results of multiple classifiers to output the detection results and classification confidence.

[0036] Specifically, the data acquisition step simultaneously acquires multi-channel signals and environmental parameters, ensuring the comprehensiveness and correlation of the data and providing a complete data foundation for subsequent processing. The dynamic environmental compensation step accurately corrects for the effects of temperature and humidity, solving the signal distortion problem caused by environmental fluctuations. The adaptive cleaning step removes outlier noise through density clustering algorithms, improving signal purity. The feature selection step eliminates redundant channels, reducing computational dimensionality and accelerating processing speed. The intelligent recognition step integrates multi-classifier results through ensemble learning, improving detection accuracy and confidence. The method of this application is highly efficient from data acquisition to result output, keeping the detection process time within 5 seconds and the false alarm rate below 2%. It also achieves non-contact detection, perfectly adapting to the screening needs of high-throughput and complex environments, providing a highly efficient, accurate, and convenient detection method for the public safety field.

[0037] In step S2, dynamic environmental compensation includes the following sub-steps: S21) Baseline Drift Correction Based on the real-time collected temperature and humidity parameters, the corresponding zero-point drift vector is retrieved from the preset database, and the zero-point drift vector is subtracted from the multi-channel response signal. S22) Sensitivity Correction The absorption suppression coefficient of water molecules on a specific wavelength band is calculated based on real-time humidity parameters, and the signal strength of a specific channel is compensated and amplified using a nonlinear gain function. The preset database contains zero-point drift vectors corresponding to multiple temperature and humidity combinations within the range of -10 to 50°C and 10% to 95% humidity. The nonlinear gain function is suitable for offsetting the suppressive effect of ambient humidity on detection sensitivity.

[0038] Specifically, the baseline drift correction step achieves precise calibration of the original signal by retrieving the zero-point drift vector from a pre-set database. This database covers a wide range of temperature and humidity combinations, ensuring accurate drift correction data even in extreme environments ranging from -10℃ to 50℃ and humidity from 10% to 95%. The sensitivity correction step effectively counteracts the suppressive effect of high temperature and humidity environments on detection sensitivity by calculating the absorbance suppression coefficient and dynamically adjusting the gain. This overcomes the shortcomings of existing compensation methods, which can only perform simple numerical reductions and cannot adapt to complex environments. Therefore, the implementation of the dynamic environmental compensation mechanism is more precise and controllable, ensuring that the system maintains high sensitivity and stability under different temperature and humidity conditions, avoiding detection errors caused by environmental changes, and further expanding the applicability of the method.

[0039] In step S3, adaptive cleaning employs the DBSCAN density clustering algorithm, including the following sub-steps: S31) Set the neighborhood radius (eps) and minimum number of samples (MinPts); S32) Traverse all data points in the high-dimensional feature space and calculate the number of neighboring points of each data point within the neighborhood radius; S33) If the number of neighboring points of a data point is greater than or equal to the minimum number of samples, then mark it as a core point and cluster it with its density-reachable neighboring points into one class; S34) If a data point is neither a core point nor in the neighborhood of any core point, then mark it as an outlier noise point and remove it from the signal sequence or perform interpolation smoothing. The neighborhood radius (eps) ranges from 0.1 to 0.5, and the minimum number of samples (MinPts) ranges from 50 to 200. The algorithm is suitable for removing interference data caused by occasional circuit noise or airflow jitter.

[0040] Specifically, this application can adapt to noise interference scenarios of different intensities, effectively eliminating outlier noise caused by circuit jitter, airflow changes, etc., and avoiding misjudging valid signals as noise.

[0041] In step S4, feature selection includes the following sub-steps: S41) During the model training phase, the Gradient Boosting Classifier is used to fit the training data; S42) Calculate the average information gain for each spectral channel during decision tree node splitting. S43) Channels with information gain below a preset threshold are defined as redundant channels and removed, while the top K channels in terms of information gain are retained as key feature channels. The value of K ranges from 5 to 15, and the cumulative contribution of information gain of key feature channels is not less than 85%, which is suitable for reducing the computational dimension of the model and improving inference efficiency.

[0042] Specifically, by using a gradient boosting classifier to calculate information gain, channels that contribute significantly to the identification of target substances can be accurately identified, ensuring that the selected key channels have strong specificity. The requirement that the cumulative contribution of information gain is not less than 85% ensures that the key channels can cover the core feature information of the target substances, avoiding the loss of effective information due to over-screening. The range of K values ​​from 5 to 15 eliminates redundant channels to reduce the computational dimensionality, while also avoiding insufficient features due to too few channels, thus improving data processing efficiency without affecting detection accuracy.

[0043] In step S5, intelligent recognition employs a Stacking ensemble learning model, including the following sub-steps: S51) Construct a Stacking ensemble learning model. The first layer base learner is deployed in parallel with at least two gradient boosting decision tree algorithm models selected from LightGBM, XGBoost or CatBoost. The second layer meta learner adopts a logistic regression or linear model. S52) Input the filtered key feature channel data into the first-layer base learner, and each base learner outputs a prediction probability vector; S53) The second-layer meta-learner receives the predicted probability vectors of each base learner and outputs the detection result and classification confidence of the target substance through weighted fusion; The first-layer base learners are LightGBM, XGBoost, and CatBoost algorithm models, and the training data for each model are divided using cross-validation to improve the stability and accuracy of the detection results.

[0044] Specifically, the first-layer base learner uses three heterogeneous gradient boosting decision tree models, leveraging the advantages of LightGBM for large-scale data processing, XGBoost for overfitting resistance, and CatBoost for class feature processing, respectively, to achieve multi-dimensional and accurate learning of feature data. The second-layer meta-learner uses weighted fusion of logistic regression or linear models to integrate the predictive advantages of each base learner and compensate for the limitations of a single model. The training data for the base learners is divided using cross-validation, which further improves the stability and generalization ability of the model and avoids overfitting.

[0045] As an additional note, the core of the dynamic environmental compensation mechanism lies in the construction and real-time application of the "temperature and humidity-baseline drift" correlation model, and the specific implementation process is as follows: 1. Establishing a correlation model between temperature / humidity and baseline drift: First, multiple environmental calibration experiments were conducted, covering the entire operating range of temperature (-10~50℃) and humidity (10%~95%). The system collected blank baseline signals (i.e., baseline values ​​without target volatiles) from the equipment under different temperature and humidity combinations. The experiment employed a controlled variable method: with a fixed humidity parameter, the temperature was adjusted at 1℃ intervals, and the drift amplitude of the baseline signal was recorded; with a fixed temperature parameter, the humidity was adjusted at 5%RH intervals, and the drift change of the baseline signal was recorded. Through fitting analysis of massive calibration data, a nonlinear correlation model between temperature / humidity and baseline drift was constructed to clarify the baseline drift patterns under different temperature and humidity conditions. For example, in calibration experiments in Kunming, when the ambient temperature was 23℃ and the humidity was 65%, the model recorded a baseline drift of 13.4%. Further quantification yielded specific correlation rules: for every 2℃ increase in temperature, the baseline drifted upward by 0.1mV (a deviation of approximately 10% from the baseline); for every 10% decrease in humidity (RH), the baseline drifted negatively by 0.05mV, providing accurate model support for real-time compensation.

[0046] 2. Synchronous acquisition of real-time data: During the actual detection process, the system collects the temperature and humidity parameters of the current detection environment in real time through the environmental sensing unit, and simultaneously acquires the original spectral signal through the multispectral sensing array to ensure that the timestamps of the temperature and humidity data are completely aligned with the original signals, laying a data foundation for accurate compensation.

[0047] 3. Calculate compensation value and correct baseline: Based on the real-time collected temperature and humidity parameters, call the preset "temperature and humidity - baseline drift" correlation model to calculate the baseline drift corresponding to the current environmental conditions. For example, if the actual temperature is 4℃ higher than the standard calibration temperature and the humidity is 10% lower than the standard calibration humidity, the total drift calculated by the model is (4 / 2)×0.1mV-0.05mV=0.15mV; then, using the method of "original signal baseline value - calculated drift", the drifted baseline is pulled back to the normal level, completing the accurate calibration of baseline drift and ensuring that the original signal is not affected by temperature and humidity fluctuations.

[0048] Example 2: Detection method based on DBSCAN and Stacking ensemble learning This embodiment details the specific algorithm flow for detecting trace target substances using the hardware system of Embodiment 1. This flow runs on the system's edge computing unit (based on an ARM Cortex-A series processor), and the specific steps are as follows: S1) Multidimensional data acquisition The system simultaneously activates the gas acquisition unit, multispectral sensor array, and environmental sensing unit, acquiring data at a sampling frequency of 10Hz to form a time-series data vector. ,in - This represents the photoelectric response signal of 32 channels, where T is the temperature parameter and H is the humidity parameter.

[0049] S2) Dynamic environmental compensation (preprocessing) To address the issue of baseline drift caused by humidity affecting the sensor, the system utilizes a built-in "temperature and humidity-baseline correction model" for dual correction: if the current ambient humidity H=85%, the system queries a preset zero-point drift database and finds that the baseline drift of channel A12 at this humidity level is ΔV=0.05V. Therefore, the corrected signal... Meanwhile, based on the effect of humidity on sensitivity suppression, the gain coefficient G is dynamically adjusted to perform nonlinear compensation on the signal amplitude, ensuring stable signal sensitivity.

[0050] S3) Adaptive Signal Cleaning (DBSCAN) To remove outliers caused by occasional circuit noise or airflow jitter, the DBSCAN density clustering algorithm is used to clean the corrected data stream: the parameters are set as follows: neighborhood radius (eps) = 0.2, minimum number of samples (MinPts) = 100; the algorithm traverses all data points in the 32-dimensional feature space and calculates the local density of each data point; isolated points with a density below the threshold are marked as environmental noise and removed to ensure the purity of the input signal for subsequent models.

[0051] S4) Key Feature Filtering The feature importance of 32 channels was calculated using a pre-trained gradient boosting classifier. Experimental results show that, for target substance detection, the channel importance is [not specified]. , , The information gain contribution of these 8 channels exceeds 85%; the system automatically selects the data from these 8 key channels to enter the next level of processing, and eliminates the remaining redundant channels, which greatly reduces the computing load.

[0052] S5) Stacking Ensemble Learning Inference A two-layer stacking ensemble model is constructed for final identification: The first layer (base learner) deploys three heterogeneous models, LightGBM, XGBoost, and CatBoost, in parallel, leveraging their respective advantages in large-scale data processing, anti-overfitting, and category feature processing to predict feature data and output their respective probability vectors; The second layer (meta learner) uses logistic regression, receiving the outputs of the three base learners as input, and outputs the final judgment result (e.g., "target substance detected") and confidence level (e.g., "98.5%)" through weighted fusion.

[0053] Comparative Example 1: Performance Comparison Experiment between the Invention and the Prior Art To clarify the technical advantages of the detection system and method of this invention, comparative experiments were conducted with existing mainstream detection technologies to verify the comprehensive performance of this invention in a simulated real-world environment, as detailed below: 1. Experimental conditions: The complex environment of public safety screening was simulated, with the temperature controlled at 30℃ and the relative humidity at 85%. Alcohol (100ppm) and perfume volatiles (20μg / m³) were mixed into the air to be tested as common environmental interfering substances. The target analytes were volatiles of target substance A (0.1μg / m³) and volatiles of target substance B (0.05μg / m³), both at trace levels, which met the low-concentration detection requirements in actual screening scenarios.

[0054] 2. Comparison of Scheme Designs: Control group (existing technology): The traditional electronic nose detection system is used. It has no dynamic environmental compensation mechanism and only uses a single SVM classifier for signal recognition. It represents the current mainstream non-contact detection technology level. Experimental group (this invention): Using the hardware system of Example 1, strictly following the detection process of Example 2, and enabling dynamic environmental compensation, DBSCAN signal cleaning and Stacking integrated learning model, the core technical solution of this invention is fully reproduced.

[0055] 3. Experimental Results and Advantages Analysis: Performance indicators Control group (existing technology) Experimental group (this invention) Technical Advantages Explanation accuracy 82.3% 96.8% This invention leverages the advantages of three heterogeneous base models through Stacking ensemble learning, addressing the underfitting problem of single models in small-sample trace detection and significantly improving recognition accuracy. False alarm rate 31.7% 1.5% The dynamic environmental compensation mechanism offsets the baseline drift effect in high humidity environments, and the DBSCAN algorithm effectively eliminates noise interference. This dual protection significantly reduces false alarms in complex environments. Response time 8.2 seconds 4.3 seconds The feature filtering step eliminates redundant channels, reducing the 32-channel data to 8 key channels, thus reducing the computational dimensionality and enabling rapid inference in conjunction with edge computing units. Detection limit 0.5μg / m³ 0.01 μg / m³ The specific band design and dynamic sensitivity compensation of the multispectral sensing array work synergistically to achieve precise capture of trace volatiles, improving the detection limit by 50 times compared to existing technologies. Therefore, the comparative experimental results show that the present invention, through the specific optimization of hardware structure and the synergistic innovation of algorithm strategy, is significantly superior to the existing technology in terms of core performance indicators such as detection sensitivity, anti-interference ability, response speed and accuracy. It effectively solves the core defects of traditional detection technology in the screening of trace substances in complex environments, such as "high number of missed detections, high false alarms and slow response", and fully meets the on-site real-time screening needs of public safety scenarios such as airports and train stations.

[0056] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-contact detection system for trace volatiles based on multispectral array sensing, characterized in that, include: A gas sampling unit is adapted to introduce an air sample from the area to be tested into the system in a non-contact manner; The detection chamber is connected to the gas collection unit, and the inner wall of the detection chamber is made of an inert material; A multispectral sensing array is disposed on the side wall of the detection cavity, and the multispectral sensing array includes several sets of photoelectric sensing units optimized for the characteristic wavelength bands of the target substance. An environmental sensing unit is disposed within the detection cavity, and the environmental sensing unit is adapted to collect temperature and humidity parameters of the detection environment in real time. A multi-band light source assembly is disposed in the detection cavity on the side away from the multispectral sensing array. The multi-band light source assembly is adapted to emit detection light covering the ultraviolet, visible and near-infrared bands. The processing and control unit is electrically connected to the gas acquisition unit, the multispectral sensing array, the environmental sensing unit, and the multi-band light source assembly, respectively. The processing and control unit is equipped with an intelligent signal processing module, which is configured to perform the following operations: Acquire the original time-series signal output by the multispectral sensing array; Using the temperature and humidity parameters, the original signal is baseline drift corrected through a preset dynamic environmental compensation model; A density-based clustering algorithm is used to clean outliers in the corrected signal, removing noisy data points with a density below a preset threshold. The cleaned feature data is input into a pre-trained ensemble learning model for inference, and the model outputs the category and concentration information of the target substance.

2. The trace volatile non-contact detection system based on multispectral array sensing according to claim 1, characterized in that, The inner wall material of the detection chamber is polytetrafluoroethylene (PTFE) or perfluoroalkoxy resin (PFA), and the air inlet and outlet of the detection chamber are streamlined.

3. The trace volatile non-contact detection system based on multispectral array sensing according to claim 1, characterized in that, The multispectral sensing array comprises 32 independent photoelectric sensing units. The response bands of the photoelectric sensing units are discontinuously distributed and specifically cover specific characteristic peak bands in the 200nm-400nm ultraviolet region, the 400nm-700nm visible light region, and the 700nm-1700nm near-infrared region.

4. The trace volatile non-contact detection system based on multispectral array sensing according to claim 1, characterized in that, The multi-band light source component is driven by pulse modulation, and the processing control unit is configured to acquire signals only during the pulse period when the light source is turned on.

5. A non-contact detection method for trace volatiles based on multispectral array sensing, characterized in that, The detection system applied to any one of claims 1-4 includes the following steps: S1) Data Acquisition The gas acquisition unit is activated to simultaneously acquire the multi-channel response signals of the multispectral sensor array and the temperature and humidity parameters of the environmental sensing unit. S2) Dynamic Environmental Compensation Based on the temperature and humidity parameters, zero-baseline correction and sensitivity nonlinearity compensation are performed on the multi-channel response signal. S3) Adaptive Cleaning The compensated multi-channel response signal is mapped to a high-dimensional feature space, and discrete noise points are identified and removed by density clustering algorithm. S4) Feature Filtering Calculate the feature importance of each spectral channel for the identification of the target substance, and select key channel combinations as input features; S5) Intelligent Recognition The filtered input features are fed into the ensemble learning model, which integrates the prediction results of multiple classifiers and outputs the detection results and classification confidence.

6. The non-contact detection method for trace volatiles based on multispectral array sensing according to claim 5, characterized in that, In step S2, dynamic environmental compensation includes the following sub-steps: S21) Baseline Drift Correction Based on the real-time collected temperature and humidity parameters, the corresponding zero-point drift vector is retrieved from the preset database, and the zero-point drift vector is subtracted from the multi-channel response signal. S22) Sensitivity Correction The absorption suppression coefficient of water molecules on a specific wavelength band is calculated based on the real-time humidity parameters, and the signal strength of a specific channel is compensated and amplified using a nonlinear gain function. The preset database contains zero-point drift vectors corresponding to multiple temperature and humidity combinations within the range of -10 to 50°C and 10% to 95% humidity. The nonlinear gain function is suitable for counteracting the suppressive effect of ambient humidity on detection sensitivity.

7. The non-contact detection method for trace volatiles based on multispectral array sensing according to claim 5, characterized in that, In step S3, adaptive cleaning employs the DBSCAN density clustering algorithm, including the following sub-steps: S31) Set the neighborhood radius (eps) and minimum number of samples (MinPts); S32) Traverse all data points in the high-dimensional feature space and calculate the number of neighboring points of each data point within the neighborhood radius; S33) If the number of neighboring points of a data point is greater than or equal to the minimum number of samples, then mark it as a core point and cluster it with its density-reachable neighboring points into one class; S34) If a data point is neither a core point nor in the neighborhood of any core point, then mark it as an outlier noise point and remove it from the signal sequence or perform interpolation smoothing. The neighborhood radius (eps) ranges from 0.1 to 0.5, the minimum number of samples (MinPts) ranges from 50 to 200, and the algorithm is suitable for removing interference data caused by occasional circuit noise or airflow jitter.

8. The non-contact detection method for trace volatiles based on multispectral array sensing according to claim 5, characterized in that, In step S4, feature selection includes the following sub-steps: S41) During the model training phase, the Gradient Boosting Classifier is used to fit the training data; S42) Calculate the average information gain for each spectral channel during decision tree node splitting. S43) Channels with information gain below a preset threshold are defined as redundant channels and removed, while the top K channels in terms of information gain are retained as key feature channels. The value of K ranges from 5 to 15, and the cumulative contribution of the information gain of the key feature channel is not less than 85%.

9. The non-contact detection method for trace volatiles based on multispectral array sensing according to claim 1, characterized in that, In step S5, intelligent recognition employs a Stacking ensemble learning model, including the following sub-steps: S51) Construct the Stacking ensemble learning model, in which at least two gradient boosting decision tree algorithm models selected from LightGBM, XGBoost or CatBoost are deployed in parallel for the first layer base learner, and the second layer meta learner adopts logistic regression or a linear model. S52) Input the filtered key feature channel data into the first layer base learner, and each base learner outputs a prediction probability vector; S53) The second-layer meta-learner receives the predicted probability vectors from each base learner and outputs the detection result and classification confidence of the target substance through weighted fusion; The first layer base learners are LightGBM, XGBoost, and CatBoost algorithm models, and the training data for each model are divided using cross-validation.