Water sample micro-plastic detection device based on fluorescent staining and AI algorithm

Abstract

The invention belongs to the technical field of environmental analysis instruments and intelligent detection, and discloses a water sample micro-plastic detection device based on fluorescent staining and an AI algorithm, an integrated micro-fluidic chip is adopted to replace a traditional pretreatment process, and online digestion and fluorescence labeling of a water sample are automatically completed through a sequential injection analysis flow path; the configured high-sensitivity fluorescence imaging module can capture feature images of the micro-plastics; a built-in lightweight multi-scale feature fusion network model can identify and count micro-plastic targets, output quantity, particle size distribution and morphological parameters in real time. According to the invention, full-process automation from sample input to result output is realized, the problems of tedious pretreatment, high manual dependency and difficulty in on-site rapid quantification in the prior art are solved, and an efficient technical means is provided for water environment micro-plastic pollution monitoring.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of environmental analysis instruments and intelligent detection technology, and particularly relates to a water sample microplastic detection device based on fluorescent staining and AI algorithm. BACKGROUND

[0002] With the large use of plastic products, microplastic (MP) pollution has become a global environmental problem. The current mainstream detection methods, such as thermal analysis-gas chromatography / mass spectrometry, can provide chemical composition information, but the equipment is expensive, the process is complex, and the physical morphological characteristics of the particles cannot be preserved. Spectral imaging methods such as micro-infrared or Raman can simultaneously obtain topography and chemical information, but the scanning speed is slow, the throughput is low, and the operating environment is high, which is difficult to meet the demand of large-scale screening.

[0003] The counting method based on optical microscope has low cost, but faces the core bottleneck: first, it relies on manual visual identification, which is low in efficiency and strong in subjectivity, and the difference between different operators is significant; second, for particles with high transparency, small size or similar optical characteristics with background impurities, the human eye is difficult to accurately identify, resulting in missed detection and misjudgment. Although some studies have tried to introduce machine learning for assistance, most algorithms lack generalization ability in complex real water sample background, and have poor segmentation accuracy for subtle targets, which cannot realize reliable automatic quantification.

[0004] Therefore, the market urgently needs a microplastic automatic analysis scheme that can run under conventional experimental conditions, has high precision, high efficiency and good economy, to support large-scale routine monitoring of environmental water.

[0005] Prior art (1) WO2021144323A1: "Detection of plastic microparticles by flow cytometry"

[0006] This patent discloses a method for analyzing plastic microparticles in water samples using spectral flow cytometry. Its features include: plastic microparticles in aqueous samples are detected by flow cytometry, and machine learning algorithms are used to identify plastic microparticles and distinguish organic matter, bacteria, minerals and other interference.

[0007] Its main features: no staining (Nile Red staining is pointed out to cause oil droplet / organic matter false signals); high-speed flow analysis is used, and machine learning is used for identification.

[0008] Compared to the proposed solution: Although machine learning is used for recognition, it does not involve the module of "sequential preprocessing of microfluidic chips (chemical digestion + density separation + fluorescence staining)," nor does it clearly define the overall structure such as image acquisition and fluorescence image segmentation and recognition, and particle size-morphology-concentration quantitative output.

[0009] Existing technology (2) CN116380854A: "A method for rapid detection of microplastic pollution in water based on fluorescence spectroscopy"

[0010] This invention proposes a method to determine the presence of microplastics in water by measuring the ratio of the emission peak values ​​of a water sample at two excitation wavelengths (235 nm and 295 nm) using a fluorescence spectrometer.

[0011] Its main features include: direct fluorescence spectroscopy measurement after sample filtration, without complex pretreatment structures; and determination of microplastic contamination by ratio.

[0012] Compared to the proposed solution: although it has the idea of ​​"fluorescence" detection, it does not use image acquisition, microfluidic chip structure, machine learning / AI image analysis, or output the functions of particle size distribution, morphological parameters, and quantity concentration.

[0013] 1. Lack of automation and integration in sample preparation: None of the above technologies integrate sample introduction, chemical digestion, density separation, and fluorescence staining into a microfluidic chip and execute them sequentially. Sample preparation is often done manually or with simple filtration, resulting in a long process that is susceptible to human error.

[0014] 2. Weak quantitative output capability for particle size, morphology, and concentration: While the above patents have detection or judgment capabilities, they do not systematically output data on the "particle size distribution," "morphological parameters," and "quantity concentration" of microplastics. Your solution, through automatic image acquisition and AI intelligent analysis modules, can directly achieve these quantitative indicators, enhancing the richness and practical value of the detection results.

[0015] 3. The combination of image acquisition and AI analysis is still rare: Existing technologies are mostly flow cytometry detection or simple fluorescence / spectral detection, lacking deep AI modules based on fluorescence image segmentation and recognition.

[0016] 4. The coupling effect of staining-separation-recognition needs to be improved: Although fluorescent staining has been proposed (such as the Nile Red method), it often encounters problems such as interference from organic staining, high background noise, and low separation efficiency. Summary of the Invention

[0017] To address the problems existing in the prior art, this invention provides a water sample microplastic detection device based on fluorescence staining and AI algorithms.

[0018] This invention is implemented as follows: a water sample microplastic detection device based on fluorescence staining and AI algorithms includes:

[0019] Sample pretreatment and staining module, automatic image acquisition module, AI intelligent analysis module;

[0020] The sample pretreatment and staining module, connected to the automatic image acquisition module, is designed with an integrated microfluidic chip and an integrated sequential injection analysis flow path to achieve a fully automated pretreatment process from sample introduction, online digestion to fluorescence staining.

[0021] An automatic image acquisition module, connected to an AI intelligent analysis module, is used to equip a high-sensitivity fluorescence imaging system capable of capturing clear fluorescence images of micron-sized microplastics;

[0022] The AI ​​intelligent analysis module, connected to the automatic image acquisition module, is used to deploy an improved deep learning segmentation model to achieve accurate identification and multi-parameter quantitative analysis of microplastics.

[0023] Furthermore, the sample pretreatment and staining module includes:

[0024] The digestion unit uses an optimized ratio of Fenton's reagent, in which... The molar ratio of H2O2 to H2O2 is 1:8-1:15, and the reaction is carried out at 40-50℃ for 30-60 minutes.

[0025] In the fluorescent staining unit, use 1-3 μg / mL Nile red ethanol solution and stain for 10-20 minutes under dark conditions;

[0026] The multi-stage filtration unit is configured with a three-stage filtration structure of 100μm, 20μm and 5μm, and integrates an ultrasonic-assisted density separation chamber.

[0027] Furthermore, the automatic image acquisition module includes:

[0028] The excitation light source uses a narrowband LED with a center wavelength of 470nm and is equipped with an excitation filter with a half-width of ≤15nm.

[0029] The optical imaging system is equipped with a long working distance objective lens with a numerical aperture of 0.6-0.8, which, together with a dichroic mirror and an emission filter, forms a highly efficient fluorescence optical path.

[0030] The image sensor uses a scientific-grade CCD with over 5 megapixels and is equipped with a semiconductor cooling device, with an operating temperature controlled between -10℃ and -20℃.

[0031] Furthermore, the deep learning model integrated in the AI ​​intelligent analysis module is an improved U-Net model, which embeds a channel-space dual attention mechanism in the skip connections of the U-Net architecture to adaptively fuse features from the encoder and decoder.

[0032] Furthermore, the improved U-Net model also includes multi-scale input branches, downsampling the original image to three scales of 512×512, 256×256 and 128×128 and inputting them into the network respectively; and introduces a deep supervision mechanism at each level of the decoder to assist the loss function in accelerating model convergence.

[0033] Furthermore, the quantitative results output by the AI ​​intelligent analysis module include: a particle size distribution histogram based on equivalent diameter, statistically classified into 0-50μm, 50-100μm, 100-500μm and >500μm; shape parameters including roundness, aspect ratio and fractal dimension; and a number concentration value converted based on the sampling volume.

[0034] The AI ​​intelligent analysis module is deployed on an embedded computing platform, using a Jetson AGX Orin processor, supporting FP16 and INT8 quantization inference, with an inference speed of over 30 frames per second; it is also equipped with a 4G / 5G communication module.

[0035] Another objective of this invention is to provide a method for detecting microplastics in water samples based on fluorescence staining and AI algorithms, comprising:

[0036] Step 1, Sample Pretreatment Stage: Sample collection, Fenton reagent digestion, density separation and Nile Red staining are completed sequentially using a sequential injection system, with the entire process taking no more than 90 minutes;

[0037] Step 2, Image Acquisition Stage: Using an autofocus system, acquire at least 10 non-overlapping fields of view with a 20× objective lens, and control the exposure time of each field of view to 100-500ms;

[0038] Step 3, Analysis and Processing Stage: Pixel-level segmentation of microplastics is achieved by improving the U-Net model, and the number of particles is counted based on connected component analysis, and particle size distribution and morphological parameters are calculated.

[0039] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the water sample microplastic detection method based on fluorescence staining and AI algorithm.

[0040] Another object of the present invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the water sample microplastic detection method based on fluorescence staining and AI algorithm.

[0041] Another objective of this invention is to provide an information data processing terminal for implementing the water sample microplastic detection device based on fluorescence staining and AI algorithms.

[0042] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0043] This invention achieves true "sample in - result out": through a highly integrated microfluidic chip and intelligent control system, the traditional pretreatment process that takes several hours is compressed to within 30 minutes. Operators only need to perform a simple sample injection operation to obtain a complete test report, which greatly improves the testing efficiency and reduces human error.

[0044] This invention overcomes the bottleneck in detecting small-sized microplastics: the entire chain of optimization, from optical imaging to algorithm analysis, increases the detection accuracy of the device for microplastics of 10-100μm from less than 60% using traditional methods to over 90%. In particular, through optimized fluorescence staining technology and background suppression algorithms, the problem of difficult identification of transparent microplastics is effectively solved.

[0045] This invention boasts strong field applicability: the device adopts a modular design, with core components limited to dimensions of 400×300×200mm and a weight not exceeding 10kg, and supports battery power. The lightweight model parameters on the edge computing platform are controlled to within 5M, achieving an inference speed of 10 frames per second on embedded devices, fully meeting the needs of rapid field detection.

[0046] It provides a wealth of morphological data: In addition to conventional quantity concentration information, the system can automatically output 16 morphological indicators based on equivalent diameter, such as particle size distribution, shape parameters (circularity, aspect ratio, fractal dimension), and surface texture characteristics, providing comprehensive data support for microplastic traceability analysis and environmental behavior research.

[0047] An intelligent quality control system has been established, featuring a built-in standard sample chamber and an automatic calibration program. System performance is automatically verified before and after each test to ensure the accuracy and comparability of test results. Simultaneously, a complete data traceability chain has been established to record data throughout the entire process, from sample collection to result output.

[0048] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0049] After the technical solution of this invention is commercialized, it is expected that R&D investment will be the main focus in the initial stage of the project (2026-2027). In 2026, R&D expenses will account for 35% of the total cost of 1.8 million yuan, primarily used for tackling core technologies and patent layout, laying the foundation for commercialization. In 2027, revenue of 1 million yuan is expected through technical services (such as enterprise-commissioned testing) and government subsidies, but the project will be temporarily in a loss-making state due to high R&D investment. With the official launch of the product in 2028, revenue will experience explosive growth, with projected sales revenue reaching 4 million yuan and net profit of 800,000 yuan for the first time, marking the project's entry into the commercial profitability stage. By 2030, the production cost ratio will gradually increase to 45% (total cost 3.2 million yuan), reflecting increased investment in raw material procurement and manufacturing process optimization after large-scale production, while sales expenses will rise for market promotion and channel construction. By 2033, with increased market penetration and product line expansion, sales revenue is expected to exceed 8 million yuan, with net profit reaching 3 million yuan, demonstrating sustained profitability and significant commercial value.

[0050] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0051] While existing research both domestically and internationally has made some progress in the field of microplastic detection—for example, Columbia University's wireless portable device achieves rapid fluorescent labeling and machine learning-based quantification—this device lacks a fully automated sample pretreatment process and has limited generalization ability for detecting microplastics in complex water samples. South China Agricultural University's HSI-CNN method is designed for soil environments and is not applicable to water samples; furthermore, hyperspectral imaging equipment is costly and difficult to popularize. This invention, through an integrated microfluidic chip design, achieves fully automated pretreatment from sample introduction and online digestion to fluorescent staining. Combined with high-sensitivity fluorescence imaging and an improved deep learning model, it overcomes the bottlenecks of low automation and poor applicability in existing technologies for different water environments, filling the technological gap in fully automated, high-precision, and low-cost on-site detection devices for microplastics in water samples both domestically and internationally.

[0052] (3) Does the technical solution of this invention solve the technical problems that people have long desired to solve but have never been able to solve successfully? Microplastic detection has long faced technical problems such as low efficiency of manual visual identification, strong subjectivity, and high false negative rate of small-sized microplastics. Although some studies have attempted to introduce machine learning assistance, most algorithms have insufficient generalization ability in complex real water sample backgrounds, poor segmentation accuracy of fine targets, and cannot achieve reliable automated quantification. This invention effectively solves the problem of difficult identification of transparent microplastics through optimized fluorescence staining process and background suppression algorithm; through improved U-Net model with channel-space dual attention mechanism and multi-scale input, the segmentation accuracy of 10-100μm microplastics is improved, and the detection accuracy is increased from less than 60% of traditional methods to more than 90%. At the same time, the device is highly integrated, portable and easy to use, and supports rapid on-site detection, solving the technical problem of rapid, accurate, and automated quantitative detection of microplastics that people have long desired to solve but have never been able to solve successfully.

[0053] (4) Does the technical solution of the present invention overcome technical bias? The present invention overcomes long-standing technical bias in this field. In the field of microplastic detection, there has long been a common perception that methods based on fluorescence staining and ordinary optical imaging, although low in cost, are easily affected by complex water sample backgrounds and are difficult to achieve high-precision and high-reliability quantitative analysis, especially when dealing with small, transparent or irregularly shaped particles, where their accuracy is far inferior to spectroscopic methods (such as FTIR and Raman). Therefore, the industry's technical development path has mainly focused on optimizing large-scale spectroscopic equipment, while generally neglecting the technical possibility of achieving low-cost, high-performance detection through deep integration of optics and algorithms.

[0054] This invention breaks through this technological prejudice through a disruptive system design. Specifically,

[0055] (1) By combining the pretreatment process of “online digestion of Fenton reagent” and “microfluidic multi-stage filtration / separation”, the interference of background organic matter is greatly eliminated from the source, creating a pure reaction environment for fluorescent staining and solving the fundamental problem of low signal-to-noise ratio in traditional fluorescence method;

[0056] (2) Through the full-chain optimization of "high-sensitivity cooled CCD imaging" and "improved U-Net model with embedded channel-space dual attention mechanism", enhanced extraction and precise segmentation of micron-scale microplastic features were achieved. The segmentation accuracy reached a level comparable to spectroscopic methods (>85%) in complex backgrounds. This proves that through innovative chemical pretreatment and advanced AI algorithm synergy, it is entirely possible to achieve near-spectral-level quantitative analysis performance on conventional optical platforms. Attached Figure Description

[0057] Figure 1This is a system block diagram of a water sample microplastic detection device based on fluorescence staining and AI algorithm provided in an embodiment of the present invention.

[0058] Figure 2 This is a flowchart of a water sample microplastic detection device based on fluorescence staining and AI algorithm provided in an embodiment of the present invention.

[0059] Figure 3 This is a block diagram of the overall structure of the device provided in the embodiment of the present invention.

[0060] Figure 4 This is a schematic diagram of the improved U-Net model structure provided in the embodiment of the present invention, with the position of the channel-space dual attention module in the skip connection specifically marked.

[0061] Figure 5 Samples of a microfiber fluorescence microscopy image dataset are displayed.

[0062] Figure 1 The module consists of: 1. Sample pretreatment and staining module; 2. Automatic image acquisition module; 3. AI intelligent analysis module. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0064] Currently, microplastic detection in the field of water environment monitoring faces two major bottlenecks: firstly, traditional microscopic observation and manual counting methods are extremely inefficient and highly subjective, making them unsuitable for large-scale water quality monitoring tasks; secondly, while conventional infrared or Raman spectroscopy can achieve high-precision identification, the equipment is expensive and the operation is complex, making it unsuitable for field or mobile scenarios. Therefore, this device achieves technological breakthroughs in three key areas—automation of the detection process, identification efficiency, and result standardization—through a fusion strategy of fluorescence staining and artificial intelligence (AI) image analysis, providing a systematic solution to the bottlenecks in the industrial application of microplastic detection.

[0065] The sample pretreatment and staining module of the device is built based on the principle of sequential injection analysis (SIA), and achieves automatic liquid distribution and reaction control through an embedded flow path in a microfluidic chip. After injection, the water sample passes sequentially through the Fenton reaction chamber, the ultrasonic separation chamber, and the staining chamber, forming a continuous fluidized reaction system. Fenton's reagent uses hydroxyl radicals to oxidize organic impurities, causing background biomass depolymerization and maximizing the exposure of the microplastic particle surface. Subsequent Nile Red staining embeds the hydrophobic regions of the plastic polymer chains in an ethanol environment, forming a stable fluorescent composite layer, providing characteristic contrast signals for subsequent optical identification.

[0066] The automatic image acquisition module achieves high signal-to-noise ratio imaging through narrow-band light-emitting diode (LED) excitation and a high numerical aperture objective lens. A dichroic mirror separates the excitation and emission light, and a charge-coupled device (CCD) suppresses dark current noise, ensuring image clarity of weakly fluorescent particles at the micrometer scale. The system automatically adjusts focus and controls exposure, generating high-resolution fluorescence image sequences during multi-field scanning, providing raw data input for subsequent algorithm analysis.

[0067] The AI ​​intelligent analysis module is based on an improved U-shaped convolutional neural network (U-Net), employing a dual attention structure to cross-fuse encoder and decoder features, achieving significant enhancement of microplastic targets from both spatial and channel dimensions. A multi-scale input strategy enables the model to represent both fine and large particles of texture, while a deep supervision mechanism accelerates training convergence and stabilizes output through hierarchical loss constraints. The binary mask output by the model is post-processed to extract connected components, achieving pixel-level particle recognition and boundary correction.

[0068] The system converts the segmentation results into equivalent circle diameter statistics, generates a particle size distribution histogram, and further extracts geometric parameters such as roundness, aspect ratio, and fractal dimension. Based on the known sampling volume and field-of-view calibration parameters, the concentration of microplastics per unit volume can be calculated. This structured data output is processed in real time by an embedded platform, and quantized into 16-bit floating-point numbers (FP16) or 8-bit integers (INT8) using FP32 to improve the analysis speed and meet the monitoring needs of on-site measurement.

[0069] This detection device forms a complete automated detection chain through the synergistic coupling of chemical pretreatment, fluorescence signal transduction, and intelligent visual recognition. Its core advantage lies in achieving near-spectral quantitative analysis of microplastics at low cost, adapting to rapid on-site screening in various aquatic environments (surface water, tap water, sewage outlets, etc.), and possessing replicable industrial application value.

[0070] like Figure 1 , Figure 3 As shown, an embodiment of the present invention provides a water sample microplastic detection device based on fluorescence staining and AI algorithm, comprising:

[0071] Sample pretreatment and staining module 1; automatic image acquisition module 2; AI intelligent analysis module 3;

[0072] The sample pretreatment and staining module 1 is connected to the image automatic acquisition module 2. It is designed with an integrated microfluidic chip and integrates a sequential injection analysis flow path to realize a fully automated pretreatment process from sample introduction, online digestion to fluorescence staining.

[0073] The automatic image acquisition module 2, connected to the AI ​​intelligent analysis module 3, is equipped with a high-sensitivity fluorescence imaging system, capable of capturing clear fluorescence images of micron-sized microplastics;

[0074] AI intelligent analysis module 3, connected to image automatic acquisition module 2, is used to deploy an improved deep learning segmentation model to achieve accurate identification and multi-parameter quantitative analysis of microplastics.

[0075] In this embodiment of the invention, the signal data processing is centered on the fluorescence signal acquired by the automatic image acquisition module 2, forming a complete analysis chain through hardware acquisition, signal conversion, algorithm recognition, and data output. First, after the microfluidic chip completes online digestion and fluorescence staining of the sample, the microplastic particles in the water sample emit fluorescence radiation signals under specific excitation wavelengths. The high-sensitivity CMOS fluorescence imaging system performs high-frame-rate imaging of the microplastic particles within the field of view, converting the spatially distributed fluorescence intensity information into a digital grayscale matrix I(x,y), and then using a 12–16 bit A / D conversion module to achieve analog-to-digital signal conversion, obtaining the original image data D0.

[0076] Subsequently, the system enters the data preprocessing stage, using Gaussian filtering and histogram equalization algorithms to denoise and correct the brightness of D0, eliminating the effects of background autofluorescence and uneven illumination. The preprocessed image D1 is then input into the improved U-Net deep learning segmentation model deployed in the AI ​​intelligent analysis module 3. This model combines residual convolutional units with an attention mechanism to perform pixel-level feature extraction and semantic segmentation on D1, outputting a binary mask M(x,y) to achieve accurate separation of microplastic particles from the background.

[0077] Next, the system extracts feature parameters for each particle based on connected component analysis and morphological operations, including projected area A, equivalent diameter d, aspect ratio r, edge fractal dimension F, and fluorescence intensity integral value I_f. Using a built-in pixel-to-scale conversion function k (obtained through microscopic calibration), the pixel units are converted into actual dimensions, achieving quantitative analysis of particle size distribution. Simultaneously, the AI ​​model clusters and classifies the particles based on the fluorescence feature weights in the training set, identifying different types of plastics (such as PE, PP, PET, etc.).

[0078] Finally, the system performs time averaging and statistical fusion on the results of multiple frames to generate a histogram of microplastic particle size distribution, a statistical table of morphological parameters, and a quantity concentration. The results are then output to the analysis terminal through a data interface, realizing fully automated processing and visualization of the detection signals.

[0079] The sample pretreatment and staining module provided in this embodiment of the invention includes:

[0080] The digestion unit uses an optimized ratio of Fenton's reagent, in which Fe... 2+ The molar ratio of H2O2 to H2O2 is 1:8-1:15, and the reaction is carried out at 40-50℃ for 30-60 minutes.

[0081] In the fluorescent staining unit, use 1-3 μg / mL Nile Red ethanol solution and stain for 10-20 minutes under dark conditions;

[0082] The multi-stage filtration unit is configured with a three-stage filtration structure of 100μm, 20μm and 5μm, and integrates an ultrasonic-assisted density separation chamber.

[0083] The automatic image acquisition module provided in this embodiment of the invention includes:

[0084] The excitation light source uses a narrowband LED with a center wavelength of 470nm and is equipped with an excitation filter with a half-width of ≤15nm.

[0085] The optical imaging system is equipped with a long working distance objective lens with a numerical aperture of 0.6-0.8, which, together with a dichroic mirror and an emission filter, forms a highly efficient fluorescence optical path.

[0086] The image sensor uses a scientific-grade CCD with over 5 megapixels and is equipped with a semiconductor cooling device, with an operating temperature controlled between -10℃ and -20℃.

[0087] like Figure 4 The deep learning model integrated in the AI ​​intelligent analysis module provided in this embodiment of the invention is an improved U-Net model. This model embeds a channel-space dual attention mechanism in the skip connections of the U-Net architecture to adaptively fuse the features of the encoder and decoder, thereby improving the segmentation accuracy of microplastic targets.

[0088] The improved U-Net model provided in this embodiment of the invention also includes a multi-scale input branch, which downsamples the original image to three scales of 512×512, 256×256 and 128×128 and inputs them into the network respectively; and introduces a deep supervision mechanism at each level of the decoder, which accelerates the convergence of the model through an auxiliary loss function and improves the small target detection capability.

[0089] The quantitative results output by the AI ​​intelligent analysis module provided in this embodiment of the invention include: a particle size distribution histogram based on equivalent diameter, statistically classified into 0-50μm, 50-100μm, 100-500μm and >500μm; shape parameters including roundness, aspect ratio and fractal dimension; and a number concentration value converted based on the sampling volume.

[0090] The AI ​​intelligent analysis module is deployed on an embedded computing platform, using a Jetson AGX Orin processor, supporting FP16 and INT8 quantization inference, with an inference speed of over 30 frames per second; it is also equipped with a 4G / 5G communication module to achieve real-time remote transmission of detection data.

[0091] like Figure 2 As shown, an embodiment of the present invention provides a method for detecting microplastics in water samples based on fluorescence staining and AI algorithms, comprising:

[0092] S101, Sample pretreatment stage: Sample collection, Fenton reagent digestion, density separation and Nile red staining are completed sequentially through a sequential injection system, with the entire process taking no more than 90 minutes.

[0093] S102, Image acquisition stage: Using an autofocus system, at least 10 non-overlapping fields of view are acquired with a 20× objective lens, and the exposure time of each field of view is controlled between 100-500ms;

[0094] S103, Analysis and Processing Stage: Pixel-level segmentation of microplastics is achieved by improving the U-Net model, and the number of particles is statistically analyzed based on connected components to calculate particle size distribution and morphological parameters.

[0095] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the water sample microplastic detection method based on fluorescence staining and AI algorithm.

[0096] Another object of the present invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the water sample microplastic detection method based on fluorescence staining and AI algorithm.

[0097] Another objective of this invention is to provide an information data processing terminal for implementing the water sample microplastic detection device based on fluorescence staining and AI algorithms.

[0098] Specific implementation of the present invention:

[0099] The specific implementation process of the sample pretreatment and staining module is as follows:

[0100] First, a 500 mL water sample was collected and injected into the pretreatment process via an automated sample introduction system. The water sample first passed through a 100 μm pore size stainless steel screen for primary filtration, removing large impurities such as leaves and sediment. It then entered a secondary filtration unit with a 20 μm pore size for further separation of medium-sized suspended solids. The entire filtration process was conducted with ultrasonic assistance at a frequency of 40 kHz and a power of 50 W, effectively preventing filter clogging and improving filtration efficiency.

[0101] The filtered water sample was fed into a 1L digestion reactor made of corrosion-resistant polytetrafluoroethylene (PTFE). Fenton's reagent was added proportionally using a precision metering pump: first, FeSO4 solution was added to achieve a final concentration of 12 mM; then, 30% H2O2 solution was slowly added to achieve a final concentration of 35 mM. The digestion process was carried out in a 45°C constant-temperature water bath, with the reaction time precisely controlled at 75 minutes. During the reaction, the sample was continuously stirred at 300 rpm using a magnetic stirrer to ensure thorough mixing of the reagent and sample. After digestion, the water sample changed from turbid to clear.

[0102] The UV-Vis spectrophotometer readings showed a decrease of more than 85% in absorbance at 254 nm, indicating that the organic matter was effectively removed.

[0103] After digestion, the sample enters a vacuum filtration system. This system uses a specially designed mixed cellulose ester membrane with a pore size of 0.45 μm and a diameter of 47 mm. The filtration device is equipped with a pressure sensor to maintain the filtration pressure within the range of -0.08 to -0.1 MPa, ensuring a stable filtration process. After filtration, a small amount of ultrapure water is used to rinse the reaction vessel and pipelines, transferring all residual microplastic particles to the surface of the filter membrane.

[0104] The microplastic-enriched filter membrane was transferred to a fully automated staining workstation. The workstation precisely sprayed a 1.5 μg / mL Nile red ethanol solution at a rate of 0.2 mL per square centimeter of filter membrane area, ensuring uniform dye coverage. The filter membrane was then transferred to a constant-temperature incubator at 25°C and incubated in the dark for 20 minutes. After incubation, the filter membrane surface was rinsed with ultrapure water at a flow rate of 10 mL per minute for 30 seconds to remove unbound dye molecules. Finally, the filter membrane was placed in a clean drying oven and dried at 35°C for 10 minutes for later use.

[0105] The operation process of the automatic image acquisition module is explained in detail below:

[0106] The processed filter membrane is placed on an automated stage, which is driven by a high-precision stepper motor with a positioning accuracy of ±2μm. The imaging system initialization program is then initiated, including: activating the temperature control system to stabilize the CCD camera's operating temperature at -15℃; calibrating the autofocus system; and initializing the stage coordinates.

[0107] A 470nm excitation source, employing a high-power LED array with a stable output power of 50mW, was activated. A narrow-band filter ensured that the excitation light's full width at half maximum (FWHM) was less than 10nm. After reflection by a dichroic mirror, the excitation light was focused onto the sample through a 20x plan-field apochromatic objective lens. The objective lens had a numerical aperture of 0.75 and a working distance of 0.95mm, ensuring the acquisition of high-resolution fluorescence images.

[0108] The fluorescence signal generated by the stimulated sample is collected by the same objective lens, passes through a dichroic mirror and a bandpass filter with a center wavelength of 550 nm (bandwidth of 25 nm), and is finally imaged on a scientific-grade CCD camera. The camera achieves a resolution of 2048×2048 pixels, a pixel size of 6.5 μm, and a quantum efficiency exceeding 80% at 550 nm. The exposure time is automatically adjusted according to the signal intensity, ranging from 100 ms to 2 s.

[0109] The system automatically executes the image acquisition procedure: first, it scans the entire filter membrane under low magnification to determine the sampling area; then, it automatically acquires at least 50 non-overlapping fields of view under a 20x objective lens, with each field of view having an image size of 665×665μm. During the acquisition process, the autofocus system operates in real time, performing fine-tuning of the focus every 5 fields of view based on an image sharpness algorithm to ensure consistent image sharpness across all images.

[0110] The specific workflow of the AI ​​intelligent analysis module is as follows:

[0111] The acquired image data is transmitted to the analysis workstation via gigabit Ethernet. The workstation is equipped with a high-performance graphics processing unit (GPU) (such as an NVIDIA RTX 4080) and has built-in proprietary analysis software developed based on the PyTorch framework.

[0112] The software first preprocesses the original images, including dark current correction, flat field correction, and background subtraction. The images are then input into an improved U-Net model for inference analysis. The model's training dataset contains 15,000 labeled images, covering microplastic samples of different materials (PE, PP, PS, PET, etc.), shapes (fibers, fragments, particles), and sizes (1-500 μm).

[0113] The core innovation of the model lies in the channel-spatial dual attention mechanism at the skip connections. The channel attention submodule first calculates global average pooling for each channel of the feature map, and then learns the weight distribution of each channel through two fully connected layers. The spatial attention submodule uses a 7×7 convolutional kernel to generate a spatial weight map, highlighting regions with significant microplastic features. The outputs of the two attention mechanisms are fused through matrix multiplication to finally generate a weighted feature map.

[0114] After the model outputs a binary segmentation mask, the post-processing algorithm performs the following operations: first, morphological opening is performed to remove noise points; then, hole filling is performed to ensure the integrity of the particle outline; finally, connected component labeling is performed to statistically analyze the geometric features of each independent region.

[0115] The analysis software automatically calculates the following parameters: particle count, equivalent circle diameter, area, circumference, roundness, and aspect ratio. Based on these parameters, the system generates a detailed test report, including: microplastic concentration (particles / L), particle size distribution based on equivalent diameter (divided into four intervals: 0-50μm, 50-100μm, 100-500μm, and >500μm), and shape classification statistics. All data is automatically saved in Excel format, and a formal test report in PDF format is also generated.

[0116] Detailed data and results from the implementation examples:

[0117] To comprehensively evaluate the device performance, we designed multiple sets of validation experiments. First, we conducted spiked recovery experiments using NIST-traceable standard polystyrene microspheres. Spiking experiments were performed at low, medium, and high concentration levels (10, 50, and 100 particles / L) for four particle size classes: 5 μm, 10 μm, 50 μm, and 100 μm. The results showed that the count recoveries for all particle size classes were between 90% and 95%, with relative standard deviations of less than 8%.

[0118] Further verification experiments were conducted using actual environmental water samples. Different types of water samples were collected, including those from rivers, lakes, and wastewater treatment plants, with each sample measured in triplicate. Comparison with traditional micro-infrared spectroscopy showed that the relative deviation of the device's detection results from the reference method was less than 15%, and the detection time was reduced from several days using traditional methods to less than 3 hours.

[0119] The detection performance was specifically evaluated for small-sized microplastics. Using fluorescently labeled microspheres of 1μm, 3μm, and 5μm, the results showed that the device achieved a 92% detection rate for particles larger than 5μm, over 80% for 3μm particles, and could also effectively identify 1μm particles.

[0120] Long-term stability tests show that after 30 days of continuous operation and daily quality control sample testing, all test results were within a controllable range, proving that the device has good stability and reliability.

[0121] Example 1: Automated Sample Pretreatment System Based on Sequential Injection Analysis

[0122] This embodiment features a dual-channel sequential injection flow path designed on an integrated microfluidic chip. The sample enters the digestion unit, separation unit, and staining unit sequentially through an inlet valve. The digestion unit is injected with Fenton's reagent at a molar ratio of 1:10, and the temperature control module maintains a constant temperature of 45 degrees Celsius for 40 minutes to remove soluble organic matter through hydroxyl radical oxidation. Subsequently, the fluid passes through an ultrasonically assisted density separation chamber, where the flow field shear intensity is controlled at a power density of 0.4 watts per square centimeter, achieving stratification of microplastics and deposited particles.

[0123] A 2 μg / mL Nile red ethanol solution was injected into the staining unit, and the reaction was carried out under closed, light-protected conditions for 15 minutes. After the reaction, particles were fractionally enriched through embedded 100 μm, 20 μm, and 5 μm three-stage filtration channels. All liquid routes were precisely controlled by a stepper motor-driven micro-injection pump, with a single processing volume of 10 mL, ensuring consistency and repeatability of sample processing.

[0124] Example 2: High-sensitivity fluorescence imaging and image acquisition system

[0125] In this embodiment, the excitation source is a narrowband LED array with a center wavelength of 470 nm, and a uniform light spot is output through an interference filter with a bandwidth of no more than 15 nm. The optical imaging system uses a long working distance objective lens with a numerical aperture of 0.75, and works in conjunction with a dichroic mirror and an emission filter to achieve efficient fluorescence channel separation.

[0126] Fluorescence images were acquired using a cooled charge-coupled device (CCD) with a resolution of 5 megapixels, operating at a temperature maintained at -15 degrees Celsius to reduce dark noise. The autofocus module achieved rapid field-of-view scanning via Z-axis step displacement, acquiring 10 non-overlapping fields of view at a time, with a fixed exposure time of 300 milliseconds. The acquired results were transmitted to an image caching server to form a standardized fluorescence dataset, providing input for subsequent algorithm training and inference.

[0127] Example 3: An Improved U-Shaped Neural Network with Embedded Dual Attention Mechanism

[0128] This embodiment addresses the issues of high background noise and large particle size differences in microplastic images by introducing dual attention modules (channel and spatial attention) on top of the standard U-shaped convolutional neural network structure. The encoder employs a residual convolutional structure to enhance feature extraction depth, while the decoder fuses high- and low-level features through skip connections. The channel attention module improves the light intensity response characteristics of plastic particles through an adaptive weighting mechanism, while the spatial attention module strengthens edge and texture feature recognition capabilities.

[0129] The network employs a multi-scale input strategy with input resolutions of 512×512, 256×256, and 128×128 pixels. The outputs of each branch are jointly optimized using a deep supervised loss function. During training, the model updates parameters using the Adam optimizer with a learning rate of 0.001 and a batch size of 8. The network achieves a microplastic recognition accuracy exceeding 95% on the test set, demonstrating its robustness and generalization performance in complex water sample images.

[0130] Example 4: Quantitative Analysis Method for Microplastic Particle Size and Morphological Parameters

[0131] This embodiment deploys a particle statistics module based on connected component labeling in the algorithm inference backend. First, the pixel area of ​​each target region is calculated based on the segmentation mask, and the particle size is calculated using the formula "equivalent diameter equals two times the area divided by pi and then taking the square root". Subsequently, three morphological indicators are extracted: circularity, aspect ratio, and fractal dimension. The formula for calculating circularity is "four times pi multiplied by the area and then divided by the square of the perimeter".

[0132] All statistical results were categorized and summarized according to particle size ranges: 0–50 μm, 50–100 μm, 100–500 μm, and greater than 500 μm. The number concentration was calculated based on sample volume and imaging field of view, and the results were output as a particle size distribution histogram and a morphological parameter table. The quantitative results were validated on multiple water samples from different bodies, showing a deviation of less than 5% from manual microscopic counting results, demonstrating the repeatability and quantitative accuracy of this method.

[0133] Example 5: Embedded AI Detection and Remote Communication System

[0134] This embodiment deploys the AI ​​intelligent analysis module on the Jetson AGX Orin embedded computing platform. The system incorporates FP16 and INT8 quantization inference engines, achieving a single-frame inference time of no more than 33 milliseconds and real-time processing of over 30 frames per second. The embedded platform connects to a remote server via Ethernet or 5G communication modules, enabling the uploading of analysis results to the cloud in encrypted JSON data structures, facilitating online data synchronization and remote monitoring.

[0135] The system adopts a modular design, including a data acquisition subsystem, an inference subsystem, and a visualization subsystem. The interface displays real-time statistics on the quantity, particle size distribution, and morphology of microplastics via a touchscreen. The overall power consumption is kept below 45 watts, allowing operation via lithium battery or external DC power supply, making it suitable for portable field testing or shipborne continuous monitoring tasks.

[0136] Figure 5 This paper presents some typical fluorescence microscopic image samples collected during the dataset construction phase of this invention. The images were acquired using the FITC channel and show the morphological characteristics of microfibers of different materials (PE, PP, PS) under fluorescence staining conditions. This reflects the diversity and clarity of the samples in the dataset and provides high-quality data support for the semantic segmentation and recognition of subsequent AI models.

[0137] To verify its technical effectiveness, embodiments of the present invention systematically constructed a high-quality image dataset for AI model training and microfiber recognition. In the dataset preparation process, three typical low-density polymers—polyethylene (PE), polypropylene (PP), and polystyrene (PS)—were selected as sample materials. These materials are not only low in density and highly hydrophobic, easy to stain, but also highly consistent with common microfiber components in the real environment, exhibiting good representativeness and practicality.

[0138] In the sample preparation stage, standard microfibers were mixed with pure water, filtered, stained with Nile Red dye, and collected by filtration through a PTFE membrane. After rinsing, they were prepared into microscopic slides for subsequent image acquisition. Image acquisition was performed on an inverted laser confocal microscope platform, using the green light channel (FITC band) as the excitation source. The obtained fluorescence signals were clear and stable with a high signal-to-noise ratio, which is beneficial for subsequent image analysis and feature extraction.

[0139] All acquired images were uniformly set to a resolution of 1024×1024 pixels to fully preserve the microfiber morphology and edge details, providing a rich information foundation for the semantic segmentation model. The sample images included both mixed fiber scenes and single-material samples to assist in subsequent annotation work and comparison and verification with the classification model.

[0140] This project collected approximately 2000 high-quality fluorescence images, and selected representative samples for pixel-level manual annotation to construct a standard semantic segmentation dataset, providing accurate supervision signals for model training. Finally, the dataset was divided into training and validation sets in an 8:2 ratio, providing a reliable basis for evaluating model training performance and optimizing the algorithm. The construction of this dataset lays a solid data foundation for verifying the technical effectiveness of the recognition method described in this invention.

[0141] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0142] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A device for detecting microplastics in water samples based on fluorescence staining and AI algorithms, characterized in that, include: The module includes a sample pretreatment and staining module, an automatic image acquisition module, and an AI intelligent analysis module. The sample pretreatment and staining module adopts an integrated microfluidic chip structure, and sequentially completes sample introduction, chemical digestion, density separation and fluorescence staining through a sequential injection analysis flow path. The automatic image acquisition module is used to acquire fluorescence images of microplastic particles under excitation light irradiation in a specific wavelength band; The AI ​​intelligent analysis module is used to segment, identify, and quantitatively calculate the fluorescence image, and output the microplastic particle size distribution, morphological parameters, and quantity concentration results.

2. The apparatus as claimed in claim 1, characterized in that, The sample pretreatment and staining module includes a digestion unit, a density separation unit, and a fluorescence staining unit. The Fenton reagent used in the digestion unit has a molar ratio of ferrous ions to hydrogen peroxide of 1:8 to 1:15, a reaction temperature of 40 to 50 degrees Celsius, and a reaction time of 30 to 60 minutes. The density separation unit is equipped with a multi-stage filtration structure, consisting of filter layers of 100 micrometers, 20 micrometers, and 5 micrometers, and the stratification effect is enhanced by ultrasonic action.

3. The apparatus as described in claim 1, characterized in that, The fluorescent staining unit uses a Nile Red ethanol solution with a concentration of 1 to 3 micrograms per milliliter, and reacts for 10 to 20 minutes under light-protected conditions. The resulting staining product is output through an optical window within the microfluidic chip to enter the imaging module.

4. The apparatus as claimed in claim 1, characterized in that, The automatic image acquisition module includes an excitation light source, an optical imaging system, and an image sensor; The excitation light source is a narrowband light-emitting diode with a center wavelength of 470 nanometers, combined with a filter with a half-width of no more than 15 nanometers; The optical imaging system employs a long working distance objective lens with a numerical aperture of 0.6 to 0.8, and is equipped with a dichroic mirror and an emission filter to form a fluorescence imaging optical path; The image sensor is a cooled charge-coupled device with a resolution of no less than 5 million pixels and an operating temperature controlled between -10 and -20 degrees Celsius.

5. The apparatus as claimed in claim 1, characterized in that, The AI ​​intelligent analysis module employs an improved U-shaped convolutional neural network structure, embedding a dual mechanism of channel attention and spatial attention in the skip connections between the encoder and decoder. This enhances the fusion capability between features at different scales and suppresses non-target background regions.

6. The apparatus as claimed in claim 5, characterized in that, The U-shaped convolutional neural network structure includes multi-scale input branches, which respectively receive images with downsampling sizes of 512×512 pixels, 256×256 pixels, and 128×128 pixels. Furthermore, auxiliary loss functions are introduced into each layer during the decoding stage to achieve deep supervision, thereby accelerating model convergence and improving segmentation accuracy.

7. A method for detecting microplastics in water samples based on the apparatus described in any one of claims 1 to 6. Its features are, Includes the following steps: Step 1, sample pretreatment stage, uses a sequential injection system to complete sample introduction, Fenton reaction digestion, ultrasonic density separation and Nile Red staining in sequence, with the entire process controlled within 90 minutes; Step 2, Image Acquisition Stage: At least 10 non-overlapping field-of-view images are captured using an autofocus system with a 20x objective lens, and the exposure time of a single frame is controlled between 100 and 500 milliseconds. Step 3, the analysis stage, uses an improved U-shaped convolutional neural network to perform pixel-level segmentation of the acquired images, and analyzes and counts the number of particles and calculates particle size and morphological parameters based on connected components.

8. The method as described in claim 7, characterized in that, In step 3, the particle size calculation uses the equivalent circle diameter method. The equivalent circle diameter is the square root of the particle area divided by pi, multiplied by two. The morphological parameters include roundness, aspect ratio, and fractal dimension. Roundness is calculated by multiplying pi by the area and then dividing by the square of the circumference.

9. An AI-powered intelligent analysis system for microplastic image recognition and analysis, characterized in that, The system consists of an input module, a feature extraction module, a segmentation module, and a quantitative calculation module; The input module receives fluorescence images and performs normalization processing; The feature extraction module uses a residual convolution structure to extract multi-scale texture features; The segmentation module performs pixel segmentation of the microplastic target region based on a channel and spatial fusion attention mechanism. The quantitative calculation module calculates the microplastic particle size distribution and morphological statistical parameters based on the segmented mask.

10. The system as described in claim 9, characterized in that, The system is deployed on an embedded computing platform and uses an inference acceleration method that supports half-precision and integer quantization to achieve real-time image analysis of at least 30 frames per second. It also uses a fourth-generation mobile communication module for remote data transmission and synchronization with the cloud.

Images