A microplastic fluorescence fingerprint detection method

Abstract

The application discloses a micro-plastic fluorescent fingerprint detection method and belongs to the technical field of fluorescent detection. In view of the problem that traditional portable fluorescent methods are easily interfered by particle size, concentration, aging degree and matrix background due to the dependence on single peak intensity signal, leading to low discrimination accuracy, at least two narrow-band LED light sources with different central wavelengths are adopted to irradiate the sample in turn or simultaneously, and the fluorescent signal intensity of at least three emission spectral windows is detected synchronously or sequentially, the intensity ratio between signals is calculated to construct a feature, and the feature is input into a pre-trained classification model to realize high-accuracy discrimination of micro-plastic polymer types. The method is mainly used for rapid identification and classification of micro-plastics in environmental water.

Description

Technical Field

[0001] This invention belongs to the field of fluorescence detection technology. Specifically, this invention relates to a method for detecting fluorescent fingerprints in microplastics. Background Technology

[0002] With the mass production and use of plastic products, plastic pollution in the environment is becoming increasingly serious. Among them, microplastics, which are tiny in size and widely distributed, have attracted much attention due to their potential ecological and health risks. Rapid identification and classification of microplastics in environmental water bodies is an important prerequisite for assessing their pollution sources, migration patterns, and potential impacts.

[0003] Currently, standard laboratory analytical methods for microplastics, such as micro-Fourier transform infrared spectroscopy and micro-Raman spectroscopy, while providing accurate compositional information, typically rely on large instruments, require complex sample pretreatment, and are time-consuming, making them unsuitable for rapid on-site screening. Therefore, developing portable and rapid on-site detection technologies is of great significance.

[0004] Fluorescence-based detection methods, due to their speed and sensitivity, show potential for portability. Existing technologies, such as the scheme disclosed in Chinese patent document CN112782141B, utilize multiple ultraviolet light-emitting diodes to excite and combine with a fluorescence detection device with multiple detection channels for rapid sorting of waste plastic products. This type of scheme typically collects absolute fluorescence intensity values ​​at multiple excitation-emission wavelengths and distinguishes materials by constructing an intensity matrix or setting empirical thresholds. This method is effective for sorting macroscopic plastic fragments that are large in size, have strong signals, and relatively simple backgrounds.

[0005] However, when such technologies are applied to the detection of microplastics in environmental water bodies, a series of prominent technical challenges arise: First, the intrinsic fluorescence signal of microplastic particles is usually extremely weak, especially without the addition of exogenous dyes, with signal intensity far lower than that of macroscopic samples, making it difficult to be stably captured by simple photoelectric systems. Second, the composition of actual environmental water samples is complex, with dissolved organic matter, algae, and their degradation products (such as humic acid) generating strong broadband fluorescence background and scattered light, severely interfering with or even drowning out the target signal. Furthermore, the concentration, particle size distribution, surface morphology, and varying degrees of environmental aging (such as photo-oxidation and thermal oxidation) of microplastics can all lead to significant changes in their fluorescence characteristics, making traditional methods relying on absolute fluorescence intensity for discrimination less robust and less accurate. The method disclosed in Chinese patent document CN112782141B, which directly acquires and compares the absolute intensities of multiple channels, demonstrates insufficient discrimination ability when facing the challenges of weak signals, strong interference, and high variability unique to microplastic detection, easily leading to misjudgments or missed detections.

[0006] In conclusion, developing a portable fluorescence detection method that can adapt to complex aquatic environments, effectively overcome background interference, and possess high sensitivity and accuracy in detecting weak signals and state changes of microplastics remains a pressing technical challenge. This requires systematic innovation and improvement at multiple levels, including the targeted selection of excitation and detection wavelengths, optimization of signal acquisition modes, effective extraction of anti-interference features, and the construction of intelligent discrimination models. Summary of the Invention

[0007] One objective of this invention is to address the technical problem of rapidly and accurately identifying the polymer type of microplastic particles with low concentration, weak signal, and severe background interference in complex aquatic environments. Microplastics themselves exhibit extremely weak fluorescence signals, making stable detection difficult using traditional portable fluorescence methods. Dissolved organic matter and suspended solids in the aquatic environment generate strong background fluorescence and scattered light, severely obscuring or distorting the target signal. Simultaneously, changes in the concentration, particle size, and aging state of microplastics can cause significant fluctuations in absolute signal intensity, rendering discrimination methods based on absolute intensity ineffective. The fluorescence spectra of different polymers (such as PS, PP, and PE) overlap, especially with the redshift after aging, making reliable differentiation difficult using simple thresholding or matrix comparison methods based on limited wavelength information.

[0008] To achieve the above objectives, this invention provides a method for detecting microplastic fluorescent fingerprints, used for rapid on-site identification of microplastic particles in environmental water bodies, comprising the following steps:

[0009] Step 1, Optical excitation: Use three narrowband LED light sources with center wavelengths of 400-410 nm (first wavelength), 455-465 nm (second wavelength), and 520-530 nm (third wavelength) to irradiate the microplastic sample to be tested in a preset time sequence.

[0010] Step 2, Multi-channel signal acquisition: Under the excitation of three LED light sources, the fluorescence signal intensity of three characteristic emission spectral windows is acquired synchronously or sequentially. The three emission spectral windows are as follows: the first emission window corresponding to the first excitation wavelength, with a center wavelength in the range of 515-525 nm; the second emission window corresponding to the second excitation wavelength, with a center wavelength in the range of 555-565 nm; and the third emission window corresponding to the third excitation wavelength, with a center wavelength in the range of 610-620 nm. In this way, multidimensional discrete fluorescence fingerprint data reflecting the polymer type and aging state of different microplastic particles are obtained.

[0011] Step 3, Robust Feature Construction: Based on the fluorescence signal intensity obtained in Step 2, calculate the intensity ratio between at least two signals from different emission windows, and construct at least one intensity ratio feature;

[0012] Step 4, Intelligent Classification and Discrimination: The intensity ratio feature constructed in Step 3 is input into a pre-trained machine learning classification model. The machine learning classification model outputs the discrimination result of the polymer type of microplastic particles in the sample, realizing the differentiation and identification of microplastics including polystyrene, polypropylene and polyethylene.

[0013] Furthermore, in step 1, the center wavelengths of the three narrowband LED light sources are 405 nm, 465 nm, and 525 nm, respectively; in step 2, the center wavelengths of the three emission spectral windows are 520 nm, 560 nm, and 615 nm, respectively. This scheme addresses the problem of ensuring that the selected wavelength combination effectively covers and distinguishes the key fingerprint features of the three core polymers, PS, PP, and PE, without using a spectrometer.

[0014] Further, the intensity ratio feature in step 3 is generated through a hierarchical feature construction process, specifically including: Step 31, based on the fluorescence intensity data obtained in step 2, firstly calculate a set of basic ratios, including: a first basic ratio, which is the ratio of the fluorescence intensity of the second emission window to the fluorescence intensity of the first emission window at the first excitation wavelength, which is used to characterize the relative intensity relationship of the two peaks of the fluorescence emission spectrum under the excitation condition; a second basic ratio, which is the ratio of the fluorescence intensity of the third emission window to the fluorescence intensity of the second emission window at the first excitation wavelength at the second excitation wavelength, which is used to sense the trend of the fluorescence emission center shifting towards the longer wavelength direction due to changes in excitation conditions or changes in the material's own state; Step 32, combine the first basic ratio, the second basic ratio, and the selected data from step 2. The original fluorescence intensity values ​​of specific channels are combined to generate derived features that are more sensitive to subtle changes in the chemical structure or physical state of materials. These features include: a first derived feature, obtained by multiplying the first base ratio by the logarithmic function of the fluorescence intensity in the third emission window at the third excitation wavelength, used to amplify the spectral differences between different aromatic polymers and improve their discriminative power; and a second derived feature, obtained by dividing the second base ratio by the first base ratio, used to specifically respond to the oxidation and aging processes of polyolefin polymers and is more sensitive to changes in their degradation degree. In step 33, the first base ratio, the second base ratio, the first derived feature, and the second derived feature are normalized. Subsequently, the normalized features are fused to form a set of comprehensive discriminant factors, which serve as the input features for the classification model in step 4. The corresponding question is: how to construct high-dimensional features sensitive to subtle changes in material aging and oxidation from discrete intensity data to improve the model's ability to discriminate subtle differences.

[0015] Furthermore, when constructing the derived combination features, the original fluorescence intensity values ​​of the selected specific channels include: a) specific intensity values ​​for constructing the first derived feature: which are the original fluorescence intensity values ​​of the third emission window under the excitation of the third excitation wavelength; b) specific intensity values ​​for assisting in the determination of the aging state of polypropylene: which are two original fluorescence intensity values ​​generated at the third emission window by excitation of the first excitation wavelength and the second excitation wavelength, respectively; wherein, the intensity value of channel a) is selected because the response of this channel signal to aromatic polymers such as polystyrene is significantly different from its response to polyolefins, and combining it with the first basic ratio can effectively amplify the distinguishability between categories; by comparing the response differences of the same emission window under different excitation wavelengths, the sensitivity and determination stability of the aging state of polypropylene polymer can be specifically enhanced by selecting the two sets of intensity values ​​of channel b).

[0016] Furthermore, the fluorescence signal intensity acquisition in step 2 employs a time-spectral two-dimensional gated synchronous detection technique, including: configuring a periodic high-frequency modulated pulse sequence for each LED, with the pulses of each LED staggered in time; the detector for each emission window is only activated and synchronously integrated during a fixed time window after the emission of its corresponding LED pulse; for the third emission window with a center wavelength in the range of 610-620 nm, its corresponding detector uses an avalanche photodiode with internal gain, and a programmable delay activation and integration time is set within the fixed time window to optimize the capture of weak fluorescence signals and avoid excitation light scattering. The corresponding problem is: how to effectively suppress inter-channel crosstalk and excitation light scattering in a multi-channel portable system to ensure stable and high signal-to-noise ratio detection of weak fluorescence signals from microplastics.

[0017] Furthermore, prior to step 1, a sample pretreatment step is included. This pretreatment step is completed online and integrated in situ with optical detection. Specifically, it includes: continuous processing of the liquid sample using a microfluidic chip. First, the sample stream is mixed with Fenton's reagent and a digestion reaction occurs within the chip channel to degrade the organic matter in the sample. Subsequently, the digested liquid stream is passed through a membrane structure to trap and enrich the target microplastic particles on the surface of the membrane structure. After enrichment, steps 1 to 4 are immediately performed on the microplastic particles enriched on the membrane structure surface for fluorescent fingerprint detection. Corresponding question: How to streamline and reliably integrate the complex laboratory sample pretreatment (filtration, digestion, enrichment) into a rapid on-site detection process to address the complex matrix interference of real-world water samples.

[0018] Furthermore, the pre-trained machine learning classification model in step 4 is a lightweight model based on random forest or gradient boosting decision tree algorithms. The training data for the machine learning classification model includes standard data of pure polystyrene, polypropylene, and polyethylene, as well as sample data subjected to accelerated aging in the laboratory and interference data with added typical environmental matrix background. When running on an embedded device, the machine learning classification model has a confidence threshold. When the output confidence is lower than the threshold, it indicates that the result is uncertain or triggers the sample reprocessing process. Corresponding question: How to construct a lightweight, highly robust intelligent discrimination model that can maintain high discrimination accuracy and adapt to sample aging and environmental matrix variations under the limited resources of embedded devices?

[0019] Furthermore, a staining enhancement discrimination mode is set up, including: before performing step 1, mixing the sample to be tested with Nile Red staining agent and incubating for a short time; in step 1, adding excitation using a fourth LED with a center wavelength of 530±10 nm; in step 2, adding detection of fluorescence intensity in a fourth emission window with a center wavelength of 580±20 nm; in steps 3 and 4, incorporating the intensity information of the fourth channel into a feature and machine learning classification model for specific identification and quantification of polyethylene and other polymers that specifically bind to the dye. Corresponding question: How to extend the detection range of the method to polymers with extremely weak intrinsic fluorescence signals (such as virgin polyethylene) without significantly increasing system complexity, thereby improving the universality of the method.

[0020] Furthermore, after short-term incubation and before formal excitation detection, a calibration step is included: a pre-scan of the stained sample using a fourth LED at low power is performed, and the pre-scan staining signal intensity value A obtained in the fourth emission window is measured and recorded; a pre-scan of the same sample area using a second or third LED at low power is performed, and the pre-scan intrinsic fluorescence signal intensity value B obtained in the third emission window is measured and recorded; the staining quality ratio Q = A / B is calculated, and this ratio Q is compared with a preset acceptable range; if the ratio Q is within the acceptable range, the staining quality is deemed acceptable, and the formal staining signal intensity subsequently measured in the fourth emission window is normalized based on the ratio Q to correct for the impact of staining process fluctuations on the quantitative results. Corresponding question: When using the staining enhancement mode, how to avoid quantitative deviations and false positive / false negative judgments caused by staining process fluctuations, differences in dye concentration, and non-specific adsorption, ensuring the reliability and repeatability of staining results?

[0021] The present invention has at least the following beneficial effects:

[0022] 1. This invention significantly improves discrimination accuracy and anti-interference capability through multi-wavelength excitation and ratio feature construction. This invention uses at least three narrowband LED light sources with different center wavelengths for sequential excitation and simultaneously collects fluorescence signals from multiple characteristic emission windows. Robust features are constructed by calculating intensity ratios, replacing the traditional method that relies on a single absolute intensity. This technique effectively overcomes signal fluctuations caused by differences in microplastic particle size and concentration, as well as environmental aging, and significantly reduces interference from background fluorescence from dissolved organic matter in water. Experiments show that this method has high discrimination accuracy for polystyrene, polypropylene, and polyethylene, and the accuracy decreases only slightly after adding humic acid interference, demonstrating excellent stability and environmental adaptability.

[0023] 2. This invention employs time-spectral two-dimensional gated synchronous detection technology to achieve high signal-to-noise ratio and low detection limit. By configuring a high-frequency modulated pulse sequence for each LED and staggering their excitation in time, coupled with synchronous integration and acquisition by the detector within the corresponding time window, physical isolation of the excitation event in the time and spectral dimensions is achieved. Specifically targeting weak fluorescence signals, an avalanche photodiode combined with programmable delay integration is used in the long-wavelength channel to effectively suppress excitation light scattering and inter-channel crosstalk. This technology enables the system to achieve stable detection of polyethylene even in a portable architecture, significantly improving the signal-to-noise ratio and reducing the detection limit by an order of magnitude compared to conventional continuous excitation modes.

[0024] 3. The integrated microfluidic online pretreatment module of this invention enables rapid and automated on-site sample preparation. This invention integrates Fenton reagent digestion and membrane enrichment steps within a microfluidic chip, allowing for in-situ integration with optical detection. It completes the degradation of organic matter and enrichment of microplastics in the sample within approximately 5 minutes. This integrated design replaces the traditional cumbersome and time-consuming laboratory pretreatment process, significantly improving detection efficiency. Furthermore, the closed-loop operation reduces sample contamination and loss, enhancing target analyte recovery and repeatability, making it particularly suitable for field and on-site emergency detection scenarios.

[0025] 4. This invention constructs a lightweight intelligent classification model that combines high accuracy, strong robustness, and low resource consumption. It employs a lightweight model based on random forests or gradient boosting decision trees, and optimizes it through a three-stage strategy of knowledge distillation, dynamic feature selection, and safe incremental learning. This allows the model to maintain discrimination accuracy close to that of complex models even on embedded devices. The dynamic feature selection mechanism effectively resists sudden interference in the field, while the incremental learning capability enables the model to continuously adapt to environmental changes and avoid knowledge forgetting. On embedded platforms such as Raspberry Pi, the model's single inference time is less than 50 ms, and its power consumption is approximately 1.1W, achieving a balance between performance and efficiency.

[0026] 5. This invention provides a staining enhancement discrimination mode, expanding the detection range of weakly fluorescent polymers. For polymers with extremely weak intrinsic fluorescence, such as polyethylene, this invention adds a staining excitation channel and a matching fluorescence acquisition window, combined with Nile Red staining and real-time staining quality optical assessment and correction steps, thereby raising the detection limit. By calculating the staining quality ratio and performing signal normalization, quantitative deviations caused by staining process fluctuations and non-specific adsorption are effectively controlled, significantly reducing the false positive rate and making staining detection a reliable and highly repeatable on-site quantitative method.

[0027] 6. The integrated portable detection system of this invention enables rapid on-site screening from sampling to identification. This invention integrates multi-wavelength optical excitation, time-gated detection, microfluidic pretreatment, intelligent discrimination models, and optional staining enhancement functions into a portable, deployable integrated device. Field tests show that the system can quickly complete the entire process from water sample collection to type identification, with results highly consistent with laboratory microscopic infrared spectroscopy methods. This significantly improves the timeliness, ease of operation, and reliability of on-site detection, systematically solving the core problems of insufficient accuracy and cumbersome procedures in traditional portable methods in complex water bodies.

[0028] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the microplastic fluorescent fingerprint detection system of the present invention;

[0030] Figure 2 The fluorescence emission spectra of polystyrene (PS), polypropylene (PP), and polyethylene (PE) at an excitation wavelength of 405 nm are shown in the figure. The shaded area in the figure indicates the first emission window of the present invention (center wavelength 520 nm, bandwidth 10 nm).

[0031] Figure 3 The fluorescence emission spectra of polystyrene (PS), polypropylene (PP), and polyethylene (PE) at an excitation wavelength of 465 nm are shown in the figure. The shaded area in the figure indicates the second emission window of the present invention (center wavelength 560 nm, bandwidth 10 nm).

[0032] Figure 4 The fluorescence emission spectra of polystyrene (PS), polypropylene (PP), and polyethylene (PE) at an excitation wavelength of 525 nm are shown in the figure. The shaded area in the figure indicates the third emission window of the present invention (center wavelength 615 nm, bandwidth 10 nm).

[0033] Figure 5This is a scatter plot of the polymer features of the present invention.

[0034] Figure 6 shows the fluorescence emission spectra of polystyrene (PS) under excitation at 405 nm, 465 nm, and 525 nm.

[0035] Figure 7 The fluorescence emission spectra of polypropylene (PP) under excitation at 405 nm, 465 nm, and 525 nm are shown.

[0036] Figure 8 The fluorescence emission spectra of polyethylene (PE) under excitation at 405 nm, 465 nm, and 525 nm are shown.

[0037] Figure 9 The fluorescence emission spectra of nylon under excitation at 405 nm, 465 nm, and 525 nm are shown.

[0038] Figure 10 The fluorescence emission spectra of polyester (PES) under excitation at 405 nm, 465 nm, and 525 nm are shown.

[0039] Figure 11 The fluorescence emission spectra of polyurethane (PU) under excitation at 405 nm, 465 nm, and 525 nm are shown.

[0040] Figure 12 The fluorescence emission spectra of polyvinyl chloride (PVC) under excitation at 405 nm, 465 nm, and 525 nm are shown.

[0041] Figure 13 This is a schematic diagram of the microplastic fluorescent fingerprint detection system of the present invention;

[0042] In the figure, 1-processing and control module, 2-signal acquisition module, 3-filter, 4-sample carrying and processing module, 5-optical excitation module. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to examples, so that those skilled in the art can implement it based on the description.

[0044] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0045] like Figures 6-12The figures show the fluorescence emission spectra of polystyrene (PS), polypropylene (PP), polyethylene (PE), nylon, polyester (PES), polyurethane (PU), and polyvinylchloride (PVC) at three characteristic excitation wavelengths. The horizontal axis represents the emission wavelength (wavelength, nm), ranging from 450 nm to 700 nm, covering the core emission spectral window defined in this invention; the vertical axis represents the fluorescence intensity (au), where au is the unit of relative intensity. The three curves in each figure correspond to the fluorescence responses at three excitation wavelengths (405 nm, 465 nm, and 525 nm, i.e., the defined first, second, and third excitation wavelengths). This provides direct spectral evidence for the subsequent construction of robust features (basic ratios and derived features), verifies the ability of the selected excitation-emission wavelength combination to distinguish various common microplastic polymers, and supports the feasibility of the method for identifying typical microplastic polymers such as PE, nylon, and polyester.

[0046] like Figure 1-4 and Figure 13 As shown, Figure 2 The fluorescence emission spectra of polystyrene (PS), polypropylene (PP), and polyethylene (PE) under an excitation wavelength of 405 nm are shown in the figure. The shaded area in the figure marks the first emission window of the present invention (center wavelength 520 nm, bandwidth 10 nm). This window is located near the fluorescence peak and can effectively capture characteristic signals and avoid excitation light scattering. Figure 3 The fluorescence emission spectra of polystyrene (PS), polypropylene (PP), and polyethylene (PE) at an excitation wavelength of 465 nm are shown in the figure. The shaded area in the figure marks the second emission window of the present invention (center wavelength 560 nm, bandwidth 10 nm), which corresponds to the characteristic response region of different polymers after redshift. Figure 4This figure shows the fluorescence emission spectra of polystyrene (PS), polypropylene (PP), and polyethylene (PE) at an excitation wavelength of 525 nm, according to the present invention. The shaded area in the figure indicates the third emission window of the present invention (center wavelength 615 nm, bandwidth 10 nm), which has high sensitivity to the intrinsic / staining signals of aged polypropylene and polyethylene. The microplastic fluorescence fingerprint detection method of the present invention is based on a microplastic fluorescence fingerprint detection system comprising an optical excitation module 5, a signal acquisition module 2, a sample carrying and processing module 4, and a processing and control module 1. The optical excitation module comprises three independent narrowband LED light sources, whose center wavelengths can be selected as 405 nm, 465 nm, and 525 nm, respectively, and the full width at half maximum (FWHM) of each light source is approximately 10 nm. These LEDs are mounted on lamp holders with heat sinks and fixed to the optical platform by adjustable brackets, so that the excitation light can be focused on the center of the sample area, causing the sample to fluoresce. The signal acquisition module 2 comprises three independent optical detection channels. The first channel is equipped with a bandpass filter with a center wavelength of 520 nm and a bandwidth of 10 nm and a silicon photodiode. The second channel is equipped with a bandpass filter with a center wavelength of 560 nm and a bandwidth of 10 nm, and a silicon photodiode. The third channel is equipped with a bandpass filter with a center wavelength of 615 nm and a bandwidth of 10 nm, and an avalanche photodiode. The filters and detectors of each detection channel are mounted on a three-dimensionally adjustable slider, positioned laterally to the sample area to collect fluorescence. The sample carrying and processing module includes a basic unit and an optional unit. The basic unit is a four-sided transparent quartz cuvette placed on a rotatable sample stage. The optional unit is an integrated microfluidic chip used for online sample digestion and particle enrichment. The processing and control module includes a microcontroller and an embedded single-board computer. The microcontroller is responsible for generating LED pulse signals and acquiring detector voltages. The embedded computer communicates with the microcontroller via a serial port, running data analysis software and classification models.

[0047] The system operates as follows: When using the basic unit, the liquid sample to be tested is injected into a cuvette, and the sample stage rotates to disperse the particles. The microcontroller sequentially drives LEDs at 405 nm, 465 nm, and 525 nm to emit 50-millisecond pulses of light. During the illumination of each LED, its corresponding detection channel synchronously activates integration, recording the fluorescence intensity. The embedded computer receives the intensity data from the three channels and calculates a pre-set intensity ratio, such as comparing the emission intensity of 560 nm under 465 nm excitation with the emission intensity of 520 nm under 405 nm excitation, to obtain a characteristic ratio. The computer inputs one or more calculated ratio features into a pre-installed random forest classification model, and the model outputs the discrimination result for polystyrene, polypropylene, or polyethylene. When using the optional unit, the sample and reagents are first pumped into the microfluidic chip via an external syringe pump. The digestion reaction is completed within the chip, and the particles are enriched on the membrane. Subsequently, the chip is placed in the optical path to perform the above optical detection process. This system, through multi-wavelength sequential excitation and ratio analysis, can reduce the dependence on absolute signal intensity. Compared to devices that use only a single excitation wavelength and rely on an absolute intensity threshold, this system is more stable to concentration changes and environmental background interference.

[0048] In the specific implementation of signal acquisition, the processing and control module is configured to perform time-gated synchronous detection. The microcontroller generates a periodic pulse sequence to control the LEDs. In a typical cycle, the 405nm LED first illuminates for 50 milliseconds, then turns off for 1 millisecond; the 465nm LED then illuminates for 50 milliseconds, turns off for 1 millisecond, and the 525nm LED then illuminates for 50 milliseconds. The integration time window of each detector is strictly synchronized with the illumination period of its corresponding LED. For the avalanche photodiode in the third detection channel, the integration window starts 3 milliseconds after the corresponding LED is turned on and lasts for 45 milliseconds to avoid acquiring the initial strong scattered light. These timing parameters can be set through a computer software interface. This system physically separates different excitation events through staggered excitation and precisely synchronized gated integration. Compared to the mode where all detectors operate continuously throughout the measurement cycle, this method effectively suppresses excitation light scattering and inter-channel crosstalk, improving the signal-to-noise ratio.

[0049] In the specific implementation of integrated microfluidic pretreatment, the optional unit of the sample carrying and processing module is a microfluidic chip made of polydimethylsiloxane. The chip is designed with sample and reagent inlets, which are mixed through a Y-shaped channel. The mixture flows through a serpentine reaction channel and finally through a polycarbonate membrane with a pore size of 5 micrometers. The membrane area of ​​the chip is transparent. The system is equipped with a dual-channel syringe pump connected to the chip inlet via flexible tubing. During operation, the syringe pump injects water sample and Fenton's reagent into the chip at a specific flow rate. A digestion reaction occurs within the serpentine channel for approximately 4 minutes, during which microplastic particles in the water sample are trapped and enriched by the downstream membrane. After the pump stops, the enriched area of ​​the chip is placed under the excitation spot of the optical module for fluorescence detection. This implementation integrates multiple steps of digestion, filtration, and enrichment into the chip for automated completion. Compared to traditional methods that require separate filtration and heating digestion devices and multiple sample transfers, this method has a shorter process, is simpler to operate, and is more suitable for field or on-site use.

[0050] In the specific implementation of the extended dye enhancement function, the optical module of the system can be supplemented with a fourth LED with a center wavelength of 530 nm and a fourth detection channel with a center wavelength of 580 nm and a bandwidth of 40 nm. When it is necessary to detect virgin polyethylene with extremely weak fluorescence signals, Nile Red dye can be added for incubation during sample preparation. During optical detection, the system adds a step to the original three-wavelength detection sequence, namely, lighting the 530 nm LED and collecting the fluorescence intensity of the 580 nm channel. After background correction, this intensity value is used as an independent feature or combined with the original ratio feature and input into the classification model for discrimination. This implementation expands the range of detectable polymers by adding an optical channel for a specific dye.

[0051] In the specific implementation of the control architecture, the microcontroller for the processing and control module uses an STM32 series development board, responsible for high-precision timing control and analog signal acquisition. The embedded computer uses a core board based on the ARM Cortex-A core, running a Linux system and Python applications. The two communicate via UART serial port. The software on the embedded computer is responsible for sending control commands, receiving intensity data, performing feature calculations and model inference, and displaying classification results and confidence scores on a connected touchscreen. This architecture, which separates the low-level hardware control from the high-level data analysis, allows the system to ensure precise synchronization of excitations while also running relatively complex machine learning algorithms.

[0052] Example 1

[0053] This embodiment verifies the performance of the method of the present invention in identifying the type of microplastic polymer. By constructing a multi-wavelength optical detection system, the method of the present invention is compared with the traditional single-peak fluorescence method to evaluate its accuracy, anti-interference ability, and response characteristics to the aging state of materials.

[0054] The experimental materials used were standard microspheres (particle size 10-200 μm) of polystyrene (PS), polypropylene (PP), and polyethylene (PE). To simulate aging, some PP samples were irradiated with ultraviolet light (0, 24, 48, 72 hours). The microspheres were prepared into suspensions with a concentration gradient of 0.01 to 5.0 mg / L using ultrapure water. Interference samples with 10 mg / L humic acid were also prepared to simulate complex water quality.

[0055] The detection system is built using three narrowband LEDs with center wavelengths of 405 nm, 465 nm, and 525 nm. Excitation light is applied perpendicularly to the sample. Three detection channels are positioned on the side of the sample: channel one is equipped with a 520 nm bandpass filter and a silicon photodiode; channel two is equipped with a 560 nm bandpass filter and a silicon photodiode; and channel three is equipped with a 615 nm bandpass filter and an avalanche photodiode. All LEDs are sequentially pulsed and controlled by a microcontroller, with each pulse width being 50 milliseconds. Each detector synchronously integrates and acquires signals only during the illumination period of its corresponding LED.

[0056] According to the method of this invention, the intensities I1 of the 520 nm channel under 405 nm excitation, I2 of the 560 nm channel under 465 nm excitation, and I3 of the 615 nm channel under 525 nm excitation are recorded sequentially. Two basic ratio features are calculated: the first basic ratio R1 is the ratio of I2 to I1, reflecting the fluorescence redshift trend; the second basic ratio R2 is the ratio of I3 to I1. Further, the derived feature F is calculated, which is the ratio of R2 to R1. The feature vector [R1, R2, F] is input into a pre-trained random forest classification model to obtain the polymer type discrimination result. As a comparison, the traditional single-peak method only uses the absolute intensities I3 of 525 nm excitation and 615 nm detection, and sets two fixed thresholds to divide them into low, medium, and high levels to correspond to different plastic types. In the laboratory, a fluorescence spectrometer is used to acquire the complete three-dimensional spectrum and perform principal component discrimination; the result is used as the true value for judgment.

[0057] First, a comparative test of the discrimination accuracy was conducted on pure and aged samples of three typical microplastics (polystyrene PS, polypropylene PP, and polyethylene PE). All samples were blindly tested sequentially using the method of this invention and the traditional single-peak method. Each sample was measured three times, and the most frequent discrimination result was taken as the final output of the method. Based on the laboratory spectral discrimination results, the accuracy of each method for different polymers was calculated (number of correctly classified samples / total number of samples). As shown in Table 1, the method of this invention exhibits excellent comprehensive discrimination performance, with average accuracies of 98.2%, 96.5%, and 94.0% for PS, unaged PP, and PE, respectively. This indicates that the features constructed based on discrete spectral information obtained from multi-wavelength excitation and through ratio calculation can effectively capture the unique spectral fingerprints of different polymers and are insensitive to fluctuations in absolute signal intensity. The performance of the traditional single-peak method is significantly poor, especially for PE, with a discrimination accuracy of only 55.2%, close to the level of random guessing (33.3%). This is because PE has the weakest intrinsic fluorescence signal, and its single-point absolute intensity I3 is easily confused with low-concentration PP or background noise signals. Furthermore, it is highly susceptible to environmental background and instrument fluctuations, causing the discrimination strategy based on a fixed threshold to completely fail. For PS and unaged PP, the accuracy of the traditional single-peak method (80.1%, 68.4%) is also far lower than that of the method of this invention, indicating the vulnerability to relying on a single intensity information. Therefore, this embodiment employs a multi-wavelength excitation combination and utilizes the emission intensity ratio to construct robust features, fundamentally overcoming the problem of low discrimination accuracy caused by the reliance on a single, volatile absolute intensity signal in traditional portable fluorescence methods.

[0058] Table 1: Comparison of the accuracy of different detection methods for identifying microplastics

[0059]

[0060] To evaluate the robustness of the method of this invention in real-world complex aquatic environments, a background interference resistance test was further designed, focusing on the impact of humic acid (a typical representative of dissolved organic matter in water) on the discrimination accuracy. Standard humic acid was weighed, dissolved in 0.01 M NaOH solution, and diluted to a final volume to prepare a stock solution with a concentration of 1000 mg / L. This stock solution was then filtered through a 0.45 μm filter and stored at 4°C. Suspensions of PS, PP, and PE standard samples used in the accuracy test were taken, and the humic acid stock solution was added to them to bring the final humic acid concentration in the samples to 10 mg / L. This concentration simulates the typical background level of eutrophic or organically polluted water bodies. Simultaneously, samples from the corresponding batches without added humic acid were retained as pure controls. Twenty-five pure samples and 25 humic acid interference samples were prepared for each polymer (PS, PP, PE). All pure and interference samples were detected using both the method of this invention and the conventional single-peak method, following the same procedures. The accuracy reduction is calculated as follows: for each polymer, the percentage decrease in accuracy under humic acid interference is calculated relative to the accuracy under pure conditions. A smaller decrease indicates stronger anti-interference capability. Feature / signal stability assessment: For the method of this invention: the coefficients of variation (CV) of the ratio features R1 and R2 for all interfering samples are calculated and compared with pure samples. CV = (standard deviation / mean) × 100%. A lower CV value indicates less fluctuation and greater stability of the feature under interference. For the traditional single-peak method: the CV of the absolute intensity signal I3 upon which all interfering samples depend is calculated.

[0061] The changes in discrimination accuracy are shown in Table 2. After adding 10 mg / L humic acid, the discrimination accuracy of the method of this invention for PS, PP, and PE decreased by less than 5%, with PE showing the largest decrease at 5.0%, but the accuracy remained at 89.0%. This indicates that the ratio features constructed by the method of this invention have a strong resistance to the broadband fluorescence background and scattered light interference introduced by humic acid. The accuracy of the traditional single-peak method decreased, with a decrease of more than 35%. Especially for PE, the accuracy dropped from 55.2% to 3.2%, meaning that its discrimination function was almost completely lost. The ratio features R1 and R2 on which the method of this invention relies have coefficients of variation of only 6.8% and 7.5% respectively under humic acid interference, which is a very small increase compared to their coefficients of variation under pure conditions (about 5-6%), showing good stability. The absolute intensity I3 on ​​which the traditional single-peak method relies has a coefficient of variation as high as 45.3% under humic acid interference, indicating that the signal is greatly disturbed and extremely unstable.

[0062] The background fluorescence and scattered light generated by humic acid are superimposed as an additive signal on the intrinsic fluorescence signal of the microplastic. For the absolute intensity I3, this superposition directly changes its value, and the amount of superposition fluctuates drastically with factors such as humic acid concentration and optical path, leading to complete inaccuracy in discrimination based on a fixed threshold. For the intensity ratio (e.g., R1=I2 / I1), since the background interference has similar spectral characteristics and intensities in two close emission windows (e.g., 560 nm and 520 nm), it is partially canceled out or normalized in the ratio calculation. Therefore, the ratio feature is far less sensitive to background fluctuations than the absolute intensity. Humic acid interference tests fully demonstrate that the method proposed in this invention, based on constructing features using multi-wavelength intensity ratios, can effectively suppress common background interference signals through mathematical operations, thereby maintaining high accuracy and stability of the discrimination model in complex water quality backgrounds. This solves the problem of unreliable results caused by severe background interference in traditional portable fluorescence methods in field applications.

[0063] Table 2: Test Results of Resistance to Humic Acid Interference

[0064]

[0065] To further verify the ability of the fluorescence features (R1, R2, F) constructed in this invention to characterize the intrinsic differences of microplastic polymers and their environmental aging state changes, a specific statistical analysis was conducted on the obtained feature data based on the completion of classification accuracy and anti-interference tests. The analysis data came from the detection results of all pure samples in the aforementioned accuracy and anti-interference tests, totaling 75 samples (25 each of PS, unaged PP, and PE), and all aged PP series samples (0, 24, 48, and 72-hour aging points, 25 samples at each point). For each sample, based on its original fluorescence intensity data (I1, I2, I3), the following formulas were used to calculate: the first basic ratio R1 = I2 / I1, the second basic ratio R2 = I3 / I1, and the derived feature F = R2 / R1. For each polymer type (PS, PE) and each PP at each aging time point, the mean and standard deviation (SD) of the corresponding feature values ​​for all samples were calculated to characterize the central distribution and dispersion of feature values ​​under that state. The results are shown in Table 3. Significant differences exist in the eigenvalues ​​of the three pure polymers, indicating good separation in the characteristic space: the R1 and R2 values ​​of polystyrene (PS) (0.15±0.05, 0.08±0.03) are significantly lower than those of polypropylene (PP) and polyethylene (PE). This is consistent with the intrinsic spectral characteristics of PS as an aromatic polymer, which has a relatively forward fluorescence emission peak and a narrow spectral shape. The R1 and R2 values ​​of polyethylene (PE) (0.45±0.15, 0.25±0.08) fall between those of PS and unaged PP, but its derived characteristic F value (0.56±0.18) is close to that of PS (0.53±0.20) and significantly higher than that of PP (0.16). This demonstrates that combining a single ratio with the combined characteristic (F) provides a finer distinguishing ability. The systematic differences in eigenvalues ​​among different polymers validate that the selected excitation-emission wavelength combination and the constructed ratio feature of this invention can effectively capture and quantify the intrinsic fluorescence fingerprint determined by the differences in polymer chemical structure. For PP samples, the eigenvalues ​​show a regular change with UV aging time: the baseline ratio R1 is 1.82±0.20 in the unaged state, and increases monotonically with aging time to 3.98±0.35 after 72 hours of aging. This increasing trend is directly related to the overall redshift of the fluorescence emission spectrum caused by the formation of conjugated structures and the increase of chromophores during the photooxidative aging process of polymers. R1 is a sensitive indicator designed to sense this redshift trend. The baseline ratio R2 also shows a certain increase (from 0.30 to 0.60), reflecting the increase in the relative fluorescence intensity in the long-wavelength region. In contrast to the significant changes in R1 and R2, the derived feature F(R2 / R1) remains relatively stable in PP samples at all aging stages, with the mean always fluctuating slightly around 0.15 (standard deviation ±0.03~0.04).This indicates that although aging significantly alters the absolute intensity and redshift of PP fluorescence (reflected by R1 and R2), the overall shape-ratio relationship of its emission spectrum (characterized by F) remains highly stable. Therefore, the F characteristic can serve as a robust class identifier to distinguish PP from other polymers (such as PE and PS), and is not easily affected by the degree of aging.

[0066] Therefore, feature R1 is highly sensitive to the aging state of PP and can serve as a potential quantitative indicator of aging degree; while feature F is highly robust for identifying the PP category. This demonstrates that the method of this invention, through the combination of multiple features, can simultaneously achieve material type identification and internal state (aging) assessment, providing a richer analytical dimension than traditional single indicators.

[0067] Table 3: Characteristic Value Responses of Different Polymers and Aging Polypropylene

[0068]

[0069] In summary, this embodiment verifies that the method of the present invention, by setting multi-wavelength excitation, acquiring key emission window signals and constructing intensity ratio characteristics, can significantly improve the accuracy and anti-interference of microplastic type identification, and can effectively reflect the aging information of materials, thus solving the problem of insufficient performance of traditional portable single-peak fluorescence methods in such applications.

[0070] Example 2

[0071] This embodiment aims to verify the effectiveness of time-spectral two-dimensional gated synchronous detection technology in improving the system's detection sensitivity and anti-crosstalk capability. The experiment focuses on evaluating the signal-to-noise ratio, detection limit, and dynamic range performance of this technology when detecting low-concentration, weakly fluorescent microplastic signals, and compares it with conventional excitation-detection modes.

[0072] The experimental materials mainly consisted of polyethylene standard microspheres with weak intrinsic fluorescence, ranging in size from 10 to 200 micrometers, supplemented by polystyrene standard microspheres as a strong fluorescence control. Single-component polyethylene suspensions with concentration gradients of 0.001, 0.01, 0.05, 0.1, 0.5, and 1.0 mg / L were prepared using ultrapure water. Simultaneously, to simulate the coexistence of multiple components and signal crosstalk in actual samples, a mixed sample with a polystyrene concentration of 1.0 mg / L and a polyethylene concentration of 0.01 mg / L was prepared.

[0073] The experimental system is based on the optical platform described in Example 1 with key improvements to implement the time-gating technology of this invention. The center wavelengths of the three excitation LEDs remain 405 nm, 465 nm, and 525 nm. The control unit is reprogrammed to generate a unique, periodic high-frequency modulated pulse sequence for each LED. Specifically, the 405 nm LED is lit first, remains lit for 50 milliseconds, and then turns off; after a 1-millisecond delay, the 465 nm LED is lit for 50 milliseconds and then turns off; after another 1-millisecond delay, the 525 nm LED is lit for 50 milliseconds and then turns off; this constitutes one cycle, with a period of 153 milliseconds. The LED pulses are completely staggered in time, with no overlap. Each detector is turned on and performs synchronous integration only during the emission of its paired LED and within a fixed time window thereafter. In particular, for the avalanche photodiode with the 615 nm detection channel, its integration window is set to turn on 3 milliseconds after the corresponding LED pulse is turned on, and the integration duration is set to 45 milliseconds to avoid the strongest scattered light at the moment the LED is turned on.

[0074] As a comparative example, two conventional detection modes were set up for performance comparison.

[0075] Comparative Example A (Continuous Excitation-Continuous Detection Mode): This mode simulates the full-time operation common in simplified or early integrated fluorescence detection devices. Specifically, the driving mode of the three narrow-band LED light sources (405 nm, 465 nm, 525 nm) is changed from pulse modulation to DC constant current drive, allowing them to emit light continuously throughout the measurement cycle, with the total light intensity remaining essentially consistent with the average light intensity in the main experimental mode. Correspondingly, the acquisition circuits of the three detectors (silicon photodiode and avalanche photodiode) are also adjusted to continuous integration mode, i.e., continuously acquiring and accumulating all incident light signals (including target fluorescence, background fluorescence, and excitation light scattering) without time gating throughout the entire measurement process. The advantage of this mode is its simple control and lack of need for precise synchronization; however, by mixing all time and spectral dimensions of signals together, it theoretically introduces the largest inter-channel crosstalk and background noise.

[0076] Comparative Example B (Simple Timing Excitation-Continuous Detection Mode): This mode simulates an improved approach, introducing time differentiation at the excitation end, but still lacking fine control at the detection end. Specifically, it maintains the exact same LED excitation timing as the main experiment of this invention, i.e., the 405 nm, 465 nm, and 525 nm LEDs are lit sequentially according to a preset pulse width (50 ms) and interval (1 ms). However, at the detection end, the synchronization gating logic between the detector and the LED pulses is eliminated. All three detectors remain continuously on and integrate throughout the entire measurement cycle (i.e., the total duration including all LED pulses and their intervals). Therefore, at any given moment, all detectors are collecting data from the currently lit LED, residual fluorescence from previously lit LEDs (if the fluorescence lifetime is long), ambient background light, and possible electronic background noise. This mode partially separates different excitation events, but due to the detectors' constant monitoring, it is still impossible to avoid the accumulation and recording of crosstalk signals from unpaired excitation light.

[0077] First, the impact of different detection technologies on the signal-to-noise ratio of low-concentration polyethylene samples was evaluated. Standardized polyethylene (PE) monodisperse microspheres (diameter: 50±10 μm, purity >99%) were used to ensure consistent optical properties. A certain mass of PE microspheres was weighed and prepared into a 1.0 mg / L stock solution with ultrapure water, which was then ultrasonically dispersed for 30 minutes to prevent aggregation. The stock solution was serially diluted to obtain a final PE test solution with a concentration of 0.01 mg / L. This concentration is near the system's expected detection limit, effectively assessing the technology's response to weak signals. Before each test, the sample vials were gently shaken to suspend the particles and ensure homogeneity.

[0078] For each test, 3 mL of 0.01 mg / L PE sample was injected into a dedicated four-sided transparent quartz cuvette and placed on a three-dimensional adjustable sample stage, ensuring consistent optical path alignment each time. On the same hardware platform, the following three detection modes were run by switching between them via software (dark noise calibration was performed before each mode): Mode I (This invention - Time-gated synchronous detection): Interleaved pulse excitation (405 nm → 465 nm → 525 nm, each pulse width 50 ms, interval 1 ms), the detector is only activated during the corresponding LED emission period + delay integration window. Mode II (Comparative Example B - Simple timing excitation): The LED pulse timing is the same as in Mode I, but the three detectors continuously integrate throughout the entire measurement cycle. Mode III (Comparative Example A - Continuous excitation): All LEDs are changed to continuous emission, and the detectors also continuously integrate, the total optical power is comparable to the average optical power of the pulse mode.

[0079] Each mode was used to perform 10 independent measurements on the same sample consecutively. After each measurement, the sample was allowed to stand for 5 seconds and the sample stage was slightly rotated to collect signals from different areas of the sample, reducing random errors caused by uneven particle distribution. All measurements were performed in a dark room with the ambient temperature controlled at 25±1 ℃ to avoid the influence of ambient light and temperature drift.

[0080] This experiment focuses on the 615 nm emission channel (corresponding to the third detection channel, equipped with an avalanche photodiode), which is most sensitive to weak signals. Signal value (S): The fluorescence intensity reading acquired during 525 nm LED excitation (modes I and II) or the entire continuous excitation period (mode III), expressed as voltage (mV). Noise value (N): Under the same optical and circuit settings, using ultrapure water as a blank sample, the same procedure was performed 10 times, and the standard deviation of the 615 nm channel signal was taken as the system noise level. Signal-to-noise ratio (SNR) calculation formula: SNR (dB) = 20 × log 10 (Signal average / noise standard deviation). The final reported signal-to-noise ratio is the average of the SNR calculated from 10 repeated measurements.

[0081] To quantify inter-channel crosstalk, an additional control group was set up: polystyrene (PS) samples (strong fluorescence) of the same concentration were measured under the same conditions. In the 615 nm channel, the signal intensity detected during non-525 nm excitation periods (such as 405 nm or 465 nm excitation periods) was recorded; this signal represents the crosstalk signal caused by excitation at other wavelengths. The crosstalk suppression ratio (dB) was defined as: Crosstalk Suppression Ratio (dB) = 20 × log 10 (Effective signal under 525 nm excitation / Crosstalk signal under non-525 nm excitation). The higher this ratio, the stronger the system's ability to isolate non-target excitation light.

[0082] The system was allowed to warm up and stabilize for 2 minutes after each mode switch. Data from 10 repeated measurements were used to calculate the mean signal-to-noise ratio and its standard deviation to assess measurement repeatability. Paired t-tests were used to analyze the statistical differences in signal-to-noise ratio between different modes (significance level set at p < 0.05).

[0083] Table 4: Performance comparison of different detection technologies in the detection of 0.01 mg / L PE samples (n=10)

[0084]

[0085] Crosstalk suppression ratio is defined as: the logarithm of the ratio of the effective signal intensity under 525 nm LED excitation to the crosstalk signal intensity under 405 nm or 465 nm LED excitation in a 615 nm channel (20 × log...). 10(Valid signal / crosstalk signal)).

[0086] As shown in Table 4, when detecting a low-concentration polyethylene sample of 0.01 mg / L, the signal-to-noise ratio of the time-spectral two-dimensional gated synchronous detection technology used in this invention reached 52.3 dB, significantly higher than the 30.8 dB of Comparative Example B and the 19.5 dB of Comparative Example A. This indicates that by synchronously gating and integrating the high-frequency interleaved pulse excitation of the LED with the detector, excitation light scattering and inter-channel crosstalk are effectively suppressed, allowing for the clear extraction of weak fluorescence signals.

[0087] In terms of crosstalk suppression capability, the crosstalk suppression ratio of the mode of this invention is 38.2 dB, which is far superior to 14.6 dB of Comparative Example B and 3.8 dB of Comparative Example A. This proves that time-gating technology can physically isolate different excitation events and avoid interference from strong fluorescent samples (such as polystyrene) to the detection channel of weak fluorescent samples (such as polyethylene).

[0088] Secondly, to systematically evaluate and compare the quantitative detection capabilities of different detection technologies, this experiment tested polyethylene (PE) concentration gradient samples, calculated the instrument's limit of detection (LOD) and limit of quantitation (LOQ), and examined the linear dynamic range. The specific experimental design and methods are as follows: Dry polyethylene standard microspheres (particle size: 50±10 μm, density: 0.92 g / cm³) were weighed. 3 A stock suspension with a concentration of 10.0 mg / L was prepared using ultrapure water (containing 0.05% Tween-20 surfactant to prevent aggregation). After thorough dispersion by probe sonication (200 W power, 15 min, ice bath), the supernatant was collected after standing for 30 min as the stock solution. The stock solution was diluted with ultrapure water (containing 0.01% Tween-20) to the following concentration gradients using a stepwise dilution method: 0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1.0 mg / L. Three parallel samples were prepared for each concentration point to evaluate repeatability. All gradient samples were stored at 4°C in the dark, equilibrated at room temperature for 30 min before testing, and gently vortexed for 10 s to ensure uniform particle suspension.

[0089] On the same optical hardware platform, three detection modes were sequentially switched (time-gated mode of this invention, simple timing mode of Comparative Example B, and continuous excitation mode of Comparative Example A). Before formal testing in each mode, dark current correction and baseline acquisition (ultrapure water + 0.01% Tween-20) were performed. Tests were conducted sequentially from low to high concentrations. 2 mL of sample was injected into a quartz cuvette at each concentration point and placed in a fixed sample position. Each concentration was measured five times in the same mode. After each measurement, the sample area was changed (the sample stage was slightly rotated), and the cuvette was cleaned (three rinses with ultrapure water and one rinse with the sample to be tested). The raw voltage output value (mV) of the 615nm channel (corresponding to the characteristic emission window of PE under 525nm excitation) was recorded for each measurement. Simultaneously, blank sample measurements were interspersed before and after each measurement sequence to monitor the stability of system noise.

[0090] For each concentration point, the signal-to-noise ratio (SNR) is calculated using the following formula: SNR = (I sample -I blank ) / SD blank Among them, I sample I is the average of 5 measurements taken at this concentration point. blank The average value of the blank signal during the same period, SD blank The standard deviation of the blank signal (n≥10) is used. Linear regression analysis was performed on the data for each detection mode, with concentration (mg / L) on the x-axis (logarithmic scale) and the corresponding signal-to-noise ratio (SNR) on the y-axis. The limit of detection (LOD) was the concentration value corresponding to SNR = 3 on the standard curve, representing the LOD for that mode. The limit of quantitation (LOQ) was the concentration value corresponding to SNR = 10 on the standard curve, representing the LOQ for that mode. To ensure the reliability of the extrapolation, the regression analysis only used data points with SNR between 3 and 100.

[0091] Plot concentration (mg / L, linear scale) on the x-axis and the corresponding net signal intensity (I) of the 615 nm channel on the y-axis. sample -I blank Plot the dose-response curve with (R0) as the ordinate. Perform linear fitting using the least squares method. Calculate the correlation coefficient R0. 2 The continuous concentration range corresponding to ≥0.99 is defined as the linear quantitative range of this method.

[0092] The ratio of the upper limit concentration of the linear range to the LOD is on the order of magnitude (10). X The range of concentrations that can be quantified by the method is represented by (). All data are expressed as mean ± standard deviation. Comparisons of LOD, LOQ, and linear range between different models were statistically analyzed based on the results of three independent experiments (each using freshly prepared gradient samples). The slope, intercept, and R-value of the linear fit were compared.2 The quantitative sensitivity, accuracy, and linearity of different detection modes were evaluated. The results are shown in Table 5. Thanks to the high signal-to-noise ratio, the detection limit of the method of this invention reaches the sub-microgram per liter level, and the dynamic range is wider.

[0093] Table 5: Quantitative performance parameters of different detection technologies for PE detection

[0094]

[0095] Finally, the ability to suppress crosstalk and accurately identify weak targets under complex signal conditions was verified by testing a mixed sample of polystyrene and polyethylene. In the continuous excitation mode of Comparative Example A, the strong polystyrene signal generated severe crosstalk in the 615 nm channel, completely drowning out the weak polyethylene signal and making it undetectable. In Comparative Example B, crosstalk was reduced, but it still resulted in a significantly higher quantification value for polyethylene. However, under the method of this invention, the 615 nm channel only showed a significant response during the 525 nm LED excitation period, effectively isolating the polystyrene fluorescence interference from the 405 nm and 465 nm LED excitations, allowing the 0.01 mg / L polyethylene signal to be detected clearly and accurately.

[0096] In summary, this embodiment demonstrates that by employing a time-spectral two-dimensional gating technique combining high-frequency interleaved pulse excitation of LEDs with synchronous gating integration of the detector, excitation light scattering and inter-channel crosstalk in portable multi-channel systems can be effectively suppressed. This technique significantly improves the signal-to-noise ratio of weak fluorescence signals, reduces the detection limit by an order of magnitude, and broadens the quantitative dynamic range, overcoming a key technical obstacle to achieving high-sensitivity microplastic detection in field devices with simplified optical structures.

[0097] Example 3

[0098] This embodiment aims to verify the practical effect of the proposed online microfluidic pretreatment technology in environmental water sample testing. By comparing the integrated microfluidic pretreatment process with traditional manual laboratory pretreatment methods, its performance in terms of processing time, operational complexity, target analyte recovery rate, and matrix removal capability is evaluated to demonstrate the applicability of this method in rapid on-site testing scenarios.

[0099] Experimental materials included actual environmental water samples and laboratory spiked samples. Actual water samples were collected from three typical environments: urban rivers, suburban scenic lakes, and nearshore seawater. After collection, samples were immediately placed in brown glass bottles, transported under refrigeration (4°C), and all pretreatment and testing were completed within 24 hours. Simultaneously, a mixed spiked sample containing standard microspheres of polystyrene, polypropylene, and polyethylene was prepared using ultrapure water, with a concentration of 1.0 mg / L for each polymer, for recovery experiments.

[0100] The online pretreatment technology employed in this invention is implemented through an integrated microfluidic chip. The chip body is made of 2 mm thick polymethyl methacrylate (PMMA) substrate, which possesses good light transmittance, chemical stability, and ease of processing, making it suitable for the optical window requirements of fluorescence detection. The filter membrane is a 5 μm pore size polycarbonate (PC) membrane, which has a high efficiency in retaining microplastic particles and can achieve good bonding with PMMA under hot-pressing conditions. A microfluidic channel pattern is cut on the PMMA substrate using a carbon dioxide laser cutter. The channel design includes: dual inlets for sample and reagent and a subsequent Y-shaped mixing zone; a serpentine reaction channel with a length of 30 cm, a width of 500 μm, and a depth of 200 μm to extend the digestion reaction time; and a 3 mm diameter circular enrichment chamber at the end for embedding the filter membrane. Pre-cut circular polycarbonate filter membranes are precisely placed in the enrichment chamber of the lower PMMA plate; another flat PMMA plate of the same size is then placed on top as the upper cover. The plates are then hot-pressed for 10 minutes at 105°C and 0.3 MPa in a hot press (or a flatbed press with temperature control) to thermoplastically bond and seal the two PMMA plates, firmly embedding the filter membrane within the chamber to form a closed microfluidic chip. After hot pressing, the chip channels remain intact, and an optically transparent detection window is formed above the filter membrane area. During the experiment, a dual-channel syringe pump is used to simultaneously pump the collected water sample and freshly prepared Fenton reagent into the chip. The sample and reagent are mixed in a serpentine channel and reacted at 55°C for approximately 4 minutes to digest organic matter. The digested liquid flows through the filter membrane, where microplastic particles are trapped and enriched on the membrane surface. The entire pretreatment process can be completed in approximately 5 minutes. Subsequently, the chip is placed directly under the optical detection module described in Example 1 to perform fluorescent fingerprint detection on the particles enriched in the transparent window area. As a comparative example, the traditional laboratory standard pretreatment method was adopted: equal volumes of water samples were subjected to vacuum filtration, membrane transfer, Fenton's reagent was added and digested in a 60°C water bath for 60 minutes, filtration was performed again, particle transfer and resuspension were performed, and finally, the samples were tested.

[0101] First, the processing efficiency and operational characteristics of the two pretreatment methods for standard spiked samples were evaluated, and the results are shown in Table 6. The microfluidic chip processing method is significantly superior to the traditional method in terms of total time, number of steps, and ease of operation. In terms of average spike recovery, the microfluidic method is also slightly better than the traditional method, and exhibits less fluctuation in results.

[0102] Table 6: Comparison of pretreatment performance of standard spiked samples

[0103]

[0104] Next, the original water samples before treatment by both methods and the final samples after treatment (elution from the surface of the microfluidic chip enrichment membrane or resuspension from the traditional method) were taken, and their three-dimensional fluorescence spectra (Ex: 250-550 nm, Em: 300-650 nm) were scanned using a fluorescence spectrometer. The total fluorescence integral intensity in the main soluble organic matter fluorescence region (typically Ex / Em: 280-350 / 350-450 nm and 350-450 / 420-550 nm) was calculated. The background suppression rate was calculated using the formula: [1 - (intensity after treatment / intensity before treatment)] × 100%. The samples after treatment by both methods (the microfluidic chip was placed directly under the light path, and the resuspension was used in the traditional method) were subjected to the fluorescence fingerprint detection of this invention to identify and count the polymer types and quantities. Simultaneously, 10-liter parallel water samples were collected from each site. After the same pre-filtration, the samples were prepared according to the standard procedure (filtration, digestion, density separation, and filter membrane preparation) and sent to analysts for full-slice scanning and spectral identification using a micro Raman spectrometer. The results were used as a reference benchmark. The consistency between the fluorescence detection results and Raman identification results obtained from the two pretreatment methods in terms of the detected polymer types and relative abundance ranking was compared. The results are shown in Table 7. Thanks to rapid and immediate online digestion, the microfluidic method showed a significantly higher inhibition rate of total fluorescence background (>92%) than the traditional method (approximately 82-88%) when treating river and lake water rich in easily degradable organic matter. Analysis suggests that the long duration (60 minutes) in the traditional digestion step may cause some large organic molecules to condense or generate new fluorescent intermediates, while the short-time flow cytometry of the microfluidic method effectively degrades small molecules and some medium-molecular-weight organic matter, reducing such side reactions. For seawater samples, the difference in inhibition rates between the two methods narrowed, which may be related to some recalcitrant humic substances and the salt effect in the seawater matrix. As shown in Table 7, based on Raman identification results, the Kappa consistency coefficient of the microfluidic pretreatment combined with fluorescent fingerprint detection in this invention is higher than that of the traditional method in all three water samples. This indicates that the microfluidic method causes less loss of target microplastics during processing and effectively suppresses the interference of background fluorescence on fingerprint detection, thus making its discrimination results closer to authoritative spectroscopic identification results. The traditional method involves cumbersome steps and multiple transfers and filtrations, which may be the reason for increased particle loss and cross-contamination risk, as well as slightly lower consistency.

[0105] Table 7: Treatment Effects of Different Pretreatment Methods on Actual Environmental Water Samples

[0106]

[0107] Finally, the stability of the integrated pretreatment module was evaluated through continuous operation testing. After processing 10 actual water samples, the chip channel did not become clogged, and the signal-to-noise ratio attenuation for direct fluorescence detection was less than 10%, indicating that the process has good robustness and repeatability.

[0108] In summary, this embodiment demonstrates that by integrating hydrodynamic focusing, rapid digestion reaction, and membrane enrichment into a single microfluidic chip and connecting it in situ with optical detection, the traditionally cumbersome and time-consuming laboratory sample preparation process can be simplified into a rapid and integrated on-site operation. This technology not only significantly improves sample processing efficiency but also effectively reduces matrix interference through process closure and automation, thereby improving the accuracy and reliability of detection results and solving the core bottleneck problem in sample pretreatment during on-site detection.

[0109] Example 4

[0110] This embodiment aims to verify the effectiveness of the lightweight model building and optimization method in achieving high-precision and robust microplastic discrimination on embedded devices. By comparing the model employing knowledge distillation, dynamic feature selection, and secure incremental learning techniques with conventional lightweight models, complex models, and static models, its comprehensive performance is evaluated under standard test sets, sudden interference scenarios, and long-term field simulation environments.

[0111] The experimental materials and systems followed those in Examples 1 and 2, including polystyrene, polypropylene, polyethylene standard microspheres, polypropylene samples with different aging degrees, and interference samples with added humic acid. All samples were tested according to the method of this invention, generating a standard dataset containing features R1, R2, and derived feature F.

[0112] Model building, deployment, and testing strictly follow the three-stage approach described above. In the model initialization phase, complete three-dimensional spectral data of 500 samples are first collected on a laboratory server using a high-performance fluorescence spectrometer, and a deep convolutional neural network is trained as the teacher model. Subsequently, discrete feature data generated from the same batch of samples is detected using a portable device. Utilizing the soft-label probability distribution provided by the teacher model, a lightweight gradient boosting decision tree model is trained using knowledge distillation techniques as the student model. In the inference phase, the trained student model is deployed on a Raspberry Pi 4B embedded device. During model runtime, the confidence intervals of each feature of the current sample are calculated in parallel. If the half-width of a feature's confidence interval is greater than 30% of its mean, that feature is temporarily discarded, and only the remaining reliable features are used for decision-making. In the evolution phase, the system automatically archives on-site samples with a discrimination confidence score higher than 0.95. Each week, using newly accumulated samples, a resilient weight consolidation algorithm is employed to safely incrementally update the model, preventing the forgetting of old knowledge.

[0113] As comparative examples, three model schemes were set up: Comparative example C is a lightweight gradient boosting decision tree model trained with the same discrete features and conventional hard labels; Comparative example D is a simplified convolutional neural network model deployed on an embedded device that directly processes pseudo-images with multi-channel intensity values; Comparative example E is a static model trained with knowledge distillation but without any incremental learning.

[0114] First, in a controlled laboratory environment, the basic performance and resource consumption of each model on a standard test set were evaluated on an embedded device. The results are shown in Table 8. The model of this invention, after knowledge distillation, significantly outperforms conventional lightweight models in accuracy and overall performance score, and approaches the performance of complex models. At the same time, its model size, inference time, and power consumption are much lower than those of complex models, achieving a balance between performance and efficiency.

[0115] Table 8: Performance and resource consumption comparison of different models on embedded devices

[0116]

[0117] Secondly, the robustness of the dynamic feature selection mechanism under simulated sudden strong interference scenarios was tested. Two interference scenarios that could occur in the field and cause instantaneous distortion of one or more features were simulated: 1. Simulating light intensity fluctuations caused by brief voltage instability of a 405 nm LED light source, or interference from micro-bubbles or temporary obstructions in the optical path. This interference mainly affects features calculated based on I1, especially R1 (R1 = I2 / I1), which may lead to abnormally high or low R1 values. 2. Simulating sudden strong ambient stray light (such as reflected light from the field) or instantaneous electrical noise pulses on the circuit board during 525 nm excitation and 615 nm signal acquisition. This interference will affect the readings of I1 and I3 simultaneously, but its impact on the calculation of derived feature F (F = R2 / R1 = (I3 / I1) / (I2 / I1)) is particularly complex and may lead to abnormal F values. Normal samples were randomly selected from the standard test set. The above interference effect is simulated at the data level using software: 1. The original intensity I1 of the normal sample is multiplied by a random perturbation factor k1 (k1 follows a normal distribution with a mean of 1.5 or 0.5 and a standard deviation of 0.2) to generate the fluorescence intensity value I1 under simulated interference conditions. d = k1 × I1, then use I1 d 1. Recalculate R1, R2, and F to construct samples with strongly perturbed R1 features. 2. Multiply the I1 and I3 of normal samples by relevant random perturbation factors to generate data that mathematically cause the F value to deviate significantly from the normal range (e.g., exceeding the mean ± 3 times the standard deviation), but whose R1 and R2 may still be within a reasonable range when viewed individually, thus constructing samples with strongly perturbed F features.

[0118] On a Raspberry Pi device, before inference on the feature vector ([R1, R2, F]) of each input sample, the model performs the following steps in parallel: Based on the distribution of the feature on the training set, it calculates the confidence interval of each feature in the current sample in real time (using the z-score method based on the mean and standard deviation). If the current value of a feature exceeds its preset confidence interval range (in the experiment, this is set to ±30% of the mean of the feature training set as the half-width threshold), the feature is determined to be unreliable or anomalous in the current sample. Once an anomalous feature is detected, a dynamic feature selection mechanism is immediately activated: only the remaining feature subsets determined to be reliable (such as only [R2, F] or only [R1, R2]) are input into the classification model for decision-making. The classification model has already been trained during the training phase through feature subset sampling (such as random forests) or has specifically trained sub-models to handle different feature combinations, thus enabling it to accommodate such dynamic changes in input dimensions.

[0119] The model of this invention is deployed on a Raspberry Pi, enabling the aforementioned dynamic feature selection mechanism. Comparative Example C has the same structure as the model of this invention (both are lightweight GBDTs trained with knowledge distillation), but disables the dynamic feature selection mechanism, always using the complete [R1, R2, F] feature vector for inference. Comparative Example E is the same as the model of this invention (i.e., a static model obtained from the same knowledge distillation training), but also disables the dynamic feature selection mechanism. Comparative Example E is designed to separate the contribution of dynamic feature selection from the model's own performance.

[0120] Test Procedure: The generated "R1 disturbed samples," "F disturbed samples," and normal "baseline samples" were input into three models for inference, and the discrimination results were recorded. At least 100 samples were tested in each disturbance scenario. The results are shown in Table 9. On normal samples, the accuracy of the three models was basically the same, indicating that their basic performance was similar. However, when facing sudden disturbances: Thanks to the dynamic feature selection mechanism, the model of this invention can automatically identify and remove disturbed abnormal features (R1 or F), and instead rely on the remaining reliable features for decision-making. Therefore, its discrimination accuracy on R1 and F disturbed samples remained at 95.2% and 94.7%, respectively, only slightly lower than the normal situation, demonstrating extremely strong robustness and fault tolerance. Comparative models C and E, lacking feature reliability judgment, were forced to use complete feature vectors containing outliers for inference. This caused the abnormal features to mislead the classification boundary like noise, resulting in a significant decrease in discrimination performance, with accuracy dropping to 83.5% and 79.8%, respectively. The similar performance of the two models indicates that the performance decline mainly stemmed from the disturbance itself, rather than whether the model underwent incremental learning.

[0121] Table 9: Robustness test of dynamic feature selection mechanism under sudden disturbances

[0122]

[0123] Finally, a 30-day simulated field long-term monitoring experiment was conducted to verify the effectiveness of the incremental learning evolution mechanism. A small number of new environmental samples with a different distribution than the initial training set were injected into the system each week, and the results are shown in Table 10. Through incremental learning, the model of this invention continuously improves its accuracy in distinguishing new environmental samples, while maintaining its ability to retain old knowledge without decline.

[0124] Table 10: Model Performance under Long-Term Simulation Field Deployment

[0125]

[0126] In summary, this embodiment demonstrates that a lightweight discriminative model suitable for embedded devices can be constructed by combining a three-stage method of knowledge distillation, dynamic feature selection, and secure incremental learning. This model not only maintains discriminative accuracy close to that of complex models under resource-constrained conditions but also effectively resists sudden field interference, possesses the ability to adapt to environmental changes and continuously optimize performance, while protecting existing knowledge from being forgotten, thus systematically improving the intelligence level and long-term usability of portable testing equipment. Figure 5 As shown, when features R1, R2, and F are mapped to spatial coordinates (X-axis: R2 / R1, Y-axis: F / R1), polystyrene (PS), polypropylene (PP), and polyethylene (PE) exhibit obvious clustering distribution in the space formed by features R1, R2, and F, verifying the high distinguishing ability of the constructed features for polymer types.

[0127] Example 5

[0128] This embodiment aims to verify the effectiveness of staining-enhanced discrimination mode in improving the detection performance of weakly fluorescent microplastics. By introducing Nile Red staining agent, adding a fourth optical channel, and combining staining quality optical assessment and correction steps, the detection sensitivity, quantitative accuracy, resistance to complex matrix interference, and repeatability of the method for low-concentration polyethylene were evaluated.

[0129] The experimental materials mainly consisted of native polyethylene standard microspheres with extremely weak intrinsic fluorescence, ranging in size from 10 to 50 micrometers, with polystyrene standard microspheres used as a control. Polyethylene suspensions with concentration gradients of 0.001, 0.003, 0.01, 0.03, 0.1, 0.3, and 1.0 mg / L were prepared using ultrapure water. To simulate a complex matrix environment, background water samples were prepared with 10 mg / L humic acid and 5 mg / L algal extract added, and 0.01 mg / L polyethylene spiked. All samples underwent pretreatment as follows: 5 mL of sample was added to 50 μL of freshly prepared Nile Red working solution (final concentration 0.1 μg / mL), vortexed, and incubated at 40°C in the dark for 10 minutes.

[0130] Based on the optical system described in Example 1, components required for the staining-enhanced discrimination mode are added. A fourth LED with a center wavelength of 530 nm is added to the excitation source. A fourth silicon photodiode detector equipped with a bandpass filter with a center wavelength of 580 nm and a bandwidth of 40 nm is added to the detection channel. Both the hardware control and software are correspondingly expanded to support the driving and signal acquisition of this channel.

[0131] The detection was performed according to the complete steps of the staining enhancement mode of this invention. After staining incubation, staining quality was first assessed: the staining signal intensity value A of the 580 nm channel was measured using a low-power pre-scan with a fourth LED; the intrinsic fluorescence signal intensity value B of the 615 nm channel was measured using a low-power pre-scan with a 525 nm LED in the same sample area; the staining quality ratio Q = A / B was calculated. The preset acceptable range was 15 to 60. If the Q value was acceptable, it was recorded and the formal detection was initiated. During the formal detection, excitation and signal acquisition were performed sequentially at four wavelengths: 405 nm, 465 nm, 525 nm, and 530 nm, obtaining the intensity of each channel, including I_580_530. The formal staining signal was normalized and corrected using the acceptable Q value, resulting in the corrected intensity I_corr = I_580_530 / Q. The I_corr is combined with the ratio features [R1, R2, F] calculated from the original three-channel data to form an extended feature vector. This vector is then input into a lightweight gradient boosting decision tree model specifically trained for coloring patterns, and the output is the type discrimination and concentration estimation results.

[0132] As comparative examples, two modes were set up: Comparative example F is the unstained mode, which uses only the original three optical channels for detection; Comparative example G is the stained but no quality assessment mode, which performs the same staining and incubation steps, but skips the Q value assessment and correction, and directly uses the measured I_580_530 original value and the original features to merge and input into the model.

[0133] First, the quantitative performance of different modes for low-concentration polyethylene was evaluated. Using concentration gradient data, the detection limit was calculated at the concentration corresponding to a signal-to-noise ratio of 3, and the results are shown in Table 11. The detection limit of the dyeing-enhanced mode of this invention reached 0.001 mg / L for polyethylene, which is two orders of magnitude higher than the undyed mode, and maintained good linearity over a wide concentration range. Although the dyed but unevaluated mode showed improved sensitivity, its quantitative linearity, repeatability, and spiked recovery were significantly worse than the method of this invention, indicating that fluctuations in the dyeing process seriously affect the results, while the quality evaluation and correction steps effectively controlled this fluctuation.

[0134] Table 11: Comparison of quantitative properties of polyethylene by different detection modes

[0135]

[0136] Secondly, the role of the staining quality assessment procedure in controlling false positives was verified by testing spiked samples with complex matrices. Using laboratory micro-Raman spectroscopy results as the true value, the false positive rate of misclassifying background particles as polyethylene was statistically analyzed, and the results are shown in Table 12. The method of this invention, through Q-value monitoring, can identify and eliminate abnormal signals caused by non-specific dye adsorption, keeping the false positive rate at a low level.

[0137] Table 12: False positive control capability of different detection modalities in complex matrix water samples

[0138]

[0139] Finally, a 5-day repeatability experiment verified the contribution of the staining quality assessment and correction steps to ensuring the reproducibility of the results. The same batch of 0.01 mg / L polyethylene standards was stained and tested daily using freshly prepared Nile Red working solution. The results showed that although the absorbance of the staining agent itself fluctuated between days, the method of this invention, through real-time Q-value correction, kept the diurnal relative standard deviation of the final quantitative results within 7.5%, demonstrating good reproducibility.

[0140] In summary, this embodiment demonstrates that by introducing Nile Red staining and a staining-enhanced discrimination mode using a fourth optical channel, the detection sensitivity of the method for weakly fluorescent polymers such as polyethylene can be improved by two orders of magnitude. More importantly, the employed staining quality optical assessment and real-time correction steps can effectively monitor and correct signal variations caused by staining process fluctuations and non-specific adsorption, significantly reducing the risk of false positives and ensuring the accuracy and repeatability of quantitative results, making staining technology a reliable method for on-site quantitative detection.

[0141] Example 6

[0142] This embodiment aims to comprehensively verify the performance of a portable system integrating multi-wavelength excitation, time-gated detection, online microfluidic pretreatment, and intelligent discrimination model. Through field testing in real-world water environments and comparison with standard laboratory analytical methods, the system's detection accuracy, timeliness, ease of operation, and continuous operational stability under field conditions are evaluated.

[0143] Three representative locations were selected for on-site testing: downstream of urban sewage discharge outlets in inland rivers, a eutrophic suburban landscape lake, and a nearshore sea area affected by tides. At each location, the system of this invention was used for on-site testing. The specific procedure was as follows: approximately 5 liters of surface water sample were collected using a portable submersible pump and immediately introduced into the system's integrated microfluidic pretreatment module. This module automatically completed filtration, digestion, and particle enrichment, taking approximately 15 minutes. Subsequently, the system's optical detection module automatically performed multi-channel fluorescent fingerprint scanning on the particles enriched on the chip. The embedded processing unit calculated features in real time and ran a lightweight classification model, outputting polymer type identification results and particle number concentration estimates within 30 minutes. Simultaneously, 10 liters of water sample were collected in parallel at each location, refrigerated, and sent to the laboratory. After filtration, digestion, and density separation according to standard operating procedures, a micro-Fourier transform infrared spectrometer was used to scan and identify random fields of view; this result was used as the gold standard for verification.

[0144] Table 13 shows a comparative analysis of the on-site testing results and the results of the laboratory standard method. In three significantly different real-world aquatic environments, the detection results of the system of this invention for the main microplastic types showed a high degree of consistency with the laboratory method results, with Kappa coefficients all above 0.75. From sampling to obtaining results, the total time taken by the system of this invention is within 1.5 to 3 hours, while the laboratory standard method requires 3 to 5 days, demonstrating a significant time efficiency advantage.

[0145] Table 13: Comparison of Consistency between On-site Test Results and Laboratory Standard Methods

[0146]

[0147] To evaluate the continuous operational stability of the integrated system, simulated field conditions were implemented in the laboratory, allowing the system to run continuously for 8 hours. During this period, performance tests were conducted every 2 hours using the same mixed standard sample of polystyrene, polypropylene, and polyethylene, and key indicators were monitored. The results are shown in Table 14. Under prolonged continuous operation, the system's discrimination accuracy remained stable, signal strength drift was minimal, and overall operation was reliable.

[0148] Table 14: Stability Test of Integrated System After 8 Hours of Continuous Operation

[0149]

[0150] This invention is easy to operate, allowing operators to easily manage the entire process from sampling to obtaining results. The system software interface is clear and provides explicit prompts for abnormal situations such as filter membrane clogging and unqualified staining. The entire device can be stored in a standard trolley case, meeting portability requirements.

[0151] In summary, this embodiment demonstrates that the integrated portable system constructed by this invention can rapidly screen multiple microplastics in on-site water samples within hours, and its detection results show good consistency with the laboratory gold standard method. The system exhibits stable performance during continuous operation and is easy to operate, systematically solving the problems of insufficient accuracy, timeliness, and ease of use faced by traditional portable detection methods in on-site applications.

[0152] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Further modifications can be readily implemented by those skilled in the art.

Claims

1. A method for microplastic fluorescent fingerprinting, characterized in that, This method is used for rapid on-site identification of microplastic particles in environmental water bodies, and includes the following steps: Step 1, Optical excitation: Use three narrowband LED light sources with center wavelengths of 400-410 nm, 455-465 nm and 520-530 nm respectively to irradiate the microplastic sample to be tested in a preset time sequence. Step 2, Multi-channel signal acquisition: Under the excitation of three LED light sources, the fluorescence signal intensity of three characteristic emission spectral windows is acquired synchronously or sequentially. The three emission spectral windows are as follows: the first emission window corresponding to the first excitation wavelength, with a center wavelength in the range of 515-525 nm; the second emission window corresponding to the second excitation wavelength, with a center wavelength in the range of 555-565 nm; and the third emission window corresponding to the third excitation wavelength, with a center wavelength in the range of 610-620 nm. In this way, multidimensional discrete fluorescence fingerprint data reflecting the polymer type and aging state of different microplastic particles are obtained. Step 3, Robust Feature Construction: Based on the fluorescence signal intensity obtained in Step 2, calculate the intensity ratio between at least two signals from different emission windows, and construct at least one intensity ratio feature; Step 4, Intelligent Classification and Discrimination: The intensity ratio feature constructed in Step 3 is input into a pre-trained machine learning classification model. The machine learning classification model outputs the discrimination result of the polymer type of microplastic particles in the sample, realizing the differentiation and identification of microplastics including polystyrene, polypropylene and polyethylene. In step 3, the intensity ratio feature is generated through a hierarchical feature construction process, which specifically includes: Step 31: Based on the fluorescence intensity data obtained in Step 2, first calculate a set of baseline ratios, including: The first basic ratio is the ratio of the fluorescence intensity of the second emission window to the fluorescence intensity of the first emission window at the first excitation wavelength. This ratio is used to characterize the relative intensity relationship of the two peaks in the fluorescence emission spectrum at the excitation wavelength. The second basic ratio is the ratio of the fluorescence intensity of the third emission window at the second excitation wavelength to the fluorescence intensity of the second emission window at the first excitation wavelength. This ratio is used to sense the overall trend of the fluorescence emission center shifting towards longer wavelengths due to changes in the excitation wavelength or changes in the material's own state. Step 32: Combine the first basic ratio, the second basic ratio, and the original fluorescence intensity values ​​of the specific channels selected in Step 2 to generate derived features that are more sensitive to subtle changes in the chemical structure or physical state of the material, including: The first derived feature, which is obtained by multiplying the first basic ratio by the logarithmic function value of the fluorescence intensity in the third emission window at the third excitation wavelength, is used to amplify the spectral differences between different aromatic polymers and improve their distinguishability. The second derived feature, obtained by dividing the second basic ratio by the first basic ratio, is used to specifically respond to the oxidation and aging process of polyolefin polymers and is more sensitive to changes in their degree of degradation. Step 33: Normalize the first base ratio, the second base ratio, the first derived feature, and the second derived feature; then, fuse the normalized features to form a set of comprehensive discriminant factors, which will serve as the input features for the classification model in step 4.

2. The method according to claim 1, characterized in that, In step 1, the center wavelengths of the three narrowband LED light sources are 405 nm, 465 nm and 525 nm, respectively; in step 2, the center wavelengths of the three emission spectral windows are 520 nm, 560 nm and 615 nm, respectively.

3. The method according to claim 2, characterized in that, When constructing derived combinatorial features, the original fluorescence intensity values ​​of the selected specific channels include: a) A specific intensity value used to construct the first derived feature: which is the original fluorescence intensity value of the third emission window under excitation at the third excitation wavelength; b) Specific intensity values ​​used to assist in determining the aging state of polypropylene: These are the two original fluorescence intensity values ​​generated at the third emission window by excitation with the first and second excitation wavelengths, respectively.

4. The method according to claim 3, characterized in that, The fluorescence signal intensity acquisition in step 2 includes: Each LED is configured with a periodic modulation pulse sequence, with the pulses of each LED staggered in time; the detector of each emission window is turned on and synchronously integrated only during the emission of its corresponding LED pulse and within a fixed time window thereafter; for the third emission window with a center wavelength in the range of 610-620 nm, the corresponding detector is an avalanche photodiode with internal gain, and a programmable delay on-time and integration time is set in the fixed time window to optimize the capture of weak fluorescence signals and avoid excitation light scattering.

5. The method according to claim 4, characterized in that, Before step 1, a sample pretreatment step is included. This sample pretreatment step is completed online and is integrated with in-situ optical detection. Specifically, it includes: continuously processing the liquid sample through a microfluidic chip. First, the sample stream is mixed with Fenton's reagent and a digestion reaction occurs in the chip channel to degrade the organic matter in the sample. Then, the digested liquid stream is passed through a membrane structure to trap and enrich the target microplastic particles on the membrane surface. After enrichment, steps 1 to 4 are immediately performed on the microplastic particles enriched on the membrane surface for fluorescent fingerprint detection.

6. The method according to claim 4, characterized in that, The pre-trained machine learning classification model in step 4 is a lightweight model based on random forest or gradient boosting decision tree algorithms; the training data of the machine learning classification model includes standard data of pure polystyrene, polypropylene, and polyethylene, as well as sample data that has undergone accelerated aging in the laboratory and interference data with typical environmental matrix background added. When running on an embedded device, the machine learning classification model has a confidence threshold. When the output confidence is lower than the threshold, it indicates that the result is uncertain or triggers the sample reprocessing process.

7. The method according to claim 6, characterized in that, It also includes a staining enhancement discrimination mode, including: Before performing step 1, the sample to be tested is mixed with Nile Red staining agent and incubated for a short time; In step 1, a fourth LED with a center wavelength of 530±10 nm is added for excitation; In step 2, the fluorescence intensity of the fourth emission window with a detection center wavelength of 580±20 nm is increased; In steps 3 and 4, the intensity information of the fourth channel is incorporated into the feature and machine learning classification model for the specific identification and quantification of polyethylene and other polymers that specifically bind to the dye.

8. The method according to claim 7, characterized in that, Following a short incubation and before the formal excitation test, a calibration step is further included: The stained sample was pre-scanned using a fourth LED at low power. The pre-scan staining signal intensity value A obtained in the fourth emission window was measured and recorded. The same sample region was pre-scanned using a second or third LED at low power, and the intensity value B of the pre-scan intrinsic fluorescence signal obtained in the third emission window was measured and recorded. Calculate the staining quality ratio Q = A / B, and compare this ratio Q with the preset acceptable range; If the ratio Q is within the acceptable range, the dyeing quality is deemed acceptable. The intensity of the formal dyeing signal measured in the fourth emission window is then normalized based on the ratio Q to correct for the impact of dyeing process fluctuations on the quantitative results.

Images