In-situ determination system and method for pesticide residue concentration of melons and fruits

Abstract

The invention belongs to the technical field of optical detection, and discloses an in-situ measurement system and method for fruit pesticide residue concentration, the system comprises a sample pool, a sample clamp, a light source, a multispectral fusion probe, a portable spectrum tester and software communicating with the portable spectrum tester; the fusion probe is provided with a plurality of spectrum channels in a modularized manner, and the spectrum channels cover ultraviolet-visible, near-infrared, fluorescent and enhanced Raman channels. The method comprises the following steps: preparing a sample; calibrating and adjusting acquisition parameters by using a standard sample; selecting a spectrum section according to a target pesticide and implementing time-sharing or synchronous acquisition; sequentially performing baseline correction, intensity normalization and background deduction on a multiband spectrum; and inputting the obtained feature vector into a linear or nonlinear fusion model for quantification, outputting the pesticide type and concentration, and recording acquisition and processing parameters to ensure the traceability of the detection process and result.

Description

Technical Field

[0001] This invention belongs to the field of optical detection technology, specifically an in-situ determination system and method for pesticide residue concentration in fruits and vegetables. Background Technology

[0002] With the strengthening of food safety supervision, pesticide residue testing has become a routine item in agricultural product quality control. Existing testing methods are mainly divided into two categories: centralized laboratory analysis and on-site rapid testing. Laboratory analysis relies on large instruments such as chromatography-mass spectrometry, which can achieve high sensitivity and high selectivity, but the testing cycle is long, the equipment is large, and the operation and maintenance costs are high, which is not conducive to rapid on-site screening. On-site rapid testing is represented by portable test strips and immunoassay, which are simple to operate and have a fast response, but the detection limit and result repeatability are generally low, often at the ppm level, and it is difficult to consistently meet the ppb level judgment requirements.

[0003] Spectroscopy is a commonly used method for pesticide detection. Ultraviolet-visible spectrophotometry is based on electronic transition absorption and is suitable for compounds with obvious absorption characteristics. Near-infrared and infrared spectroscopy record molecular vibrational information for functional group identification. Fluorescence spectroscopy depends on the emission characteristics of the sample or target. Raman spectroscopy provides molecular fingerprint vibrational information. However, the applicability of the above single spectroscopic methods on complex matrices is limited: pigments on the sample surface, such as chlorophyll, overlap absorption with soluble components in the ultraviolet-visible region; the infrared region is interfered with by strong water absorption; fluorescence measurement is easily affected by the background fluorescence of the matrix or photobleaching effect; Raman measurement is often obscured by background fluorescence. In addition, solid or non-uniform surface samples such as fruits and vegetables need to be simplified or exempted from complex pretreatment in the field to maintain non-destructive detection. Existing methods still struggle to balance sensitivity, anti-interference and convenience in this regard. Summary of the Invention

[0004] The purpose of this invention is to provide an in-situ determination system and method for pesticide residue concentration in fruits and vegetables, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an in-situ determination system for pesticide residue concentration in fruits and vegetables, comprising: a sample cell, a sample clamp for fixing and positioning the area to be measured, a light source, a multispectral fusion probe, a portable multifunctional spectrometer, and computer software communicatively connected to the spectrometer. The multispectral fusion probe has a modular structure, including an ultraviolet-visible spectrophotometer module, a near-infrared spectrometer module, a fluorescence excitation module, and an enhanced Raman probe. The reflectance spectrum acquired by the ultraviolet-visible spectrophotometer module is recorded by a spectral recording unit and used as a baseline reference for epidermal pigmentation by the computer software. The near-infrared spectrometer module and the enhanced Raman probe achieve synchronous acquisition through optical coupling or fiber optic splitting. The raw spectra recorded by the spectral recording unit of each module are transmitted to the computer software as input for fusion quantitative analysis. Based on the above structure, this invention forms a unified detection process through a clear data flow and processing sequence: First, the ultraviolet-visible module obtains the spectral characteristics of epidermal pigments and uses them as a baseline reference. The computer software uses this reference to estimate and subtract the background components of fluorescence and Raman signals in the frequency or time domain. At the same time, the vibration absorption information collected by the near-infrared module and the fingerprint information collected by the Raman probe are input as parallel inputs into the fusion quantitative model (such as partial least squares regression or artificial neural network) within the same time window for the analysis of pesticide types and concentrations in the sample. The above process is executed according to a predetermined time sequence under the control of the spectral recording unit and software, ensuring that the causal chain from acquisition, recording to processing is clear and repeatable. This provides an feasible technical path for multi-spectral coupling determination on the surface of fruits and vegetables containing epidermal pigment background.

[0006] Furthermore, the operating or scanning range of each spectroscopic module of the multispectral fusion probe is set as follows: the operating wavelength of the ultraviolet-visible spectrophotometer module is 200–760 nm; the operating wavelength of the near-infrared spectroscopy module is 900–2500 nm; the excitation wavelength of the fluorescence spectroscopy excitation module is 280–450 nm; and the Raman shift scanning range of the enhanced Raman probe is 200–2000 cm⁻¹. -1 The above range settings are based on the complementary nature of different spectral channels for pesticide molecular structure information: In the 200–760 nm range, electronic absorption characteristics such as those containing benzene rings and conjugated double bonds can be obtained, which is beneficial for identifying pesticides with these chemical groups; in the 900–2500 nm range, the vibrational content and overtones / combination bands of C–H, O–H, and other chemical bonds can be recorded, serving as a supplement to molecular vibrational information; excitation wavelengths of 280–450 nm can induce fluorescence emission from some pesticides or their derivatives, and combined with time-resolved acquisition, transient background emission can be distinguished from target emission; 200–2000 cm⁻¹… -1As a typical Raman fingerprint region, it can acquire molecular characteristic vibrational peaks for component identification. The information obtained from each spectral band is input into the fusion model according to a predetermined time sequence and synchronization strategy during the data processing stage to form independent and complementary chemical characterization data, thereby providing clear spectroscopic basis for subsequent feature extraction and quantitative analysis.

[0007] Furthermore, the enhanced Raman probe adopts a surface-enhanced Raman scattering (SERS) substrate structure, the substrate being composed of gold nanoparticles with a particle size of approximately 50 nm, which are electrostatically self-assembled and immobilized on the surface of a silicon wafer carrier. The Raman probe is connected to the incident port of a portable multifunctional spectrometer via an optical coupler or optical fiber. The Raman signal is transmitted through the optical fiber and recorded by the spectral recording unit. The Raman fingerprint data and the spectroscopic information recorded by the near-infrared module and the ultraviolet-visible module are input into the computer software in parallel according to a predetermined time sequence as multi-spectral feature inputs for feature extraction and quantitative analysis.

[0008] Furthermore, the portable multifunctional spectrometer includes an adjustable slit, a lens, a grating, and a high-resolution CMOS imaging module arranged in sequence. The adjustable slit consists of a micrometer screw, a spring, and a pair of blades. The micrometer screw drives the blades to make a slight linear movement to change the slit width, with the initial slit width set at 50 μm. The lens consists of two biconvex lenses with a focal length of 3 cm, used to focus and collimate the reflected or scattered light from the sample into the slit and dispersion unit. The grating has a size of 20 mm × 20 mm × 2 mm and a line density of 1000 lines / mm, used to disperse the incident light into bands according to wavelength. The CMOS imaging module has approximately 5 million effective pixels and a pixel size of approximately 2.2 μm, used to record the spectral bands representing the wavelength distribution after dispersion and transmit the original spectral image to computer software. The above components are set and connected according to the above parameters. The slit width, the optical aperture of the lens, the dispersion characteristics of the grating, and the pixel size of the CMOS form a matching relationship within the spectral recording unit, so that the recorded spectral bands have a distinguishable wavelength distribution on the CMOS, which facilitates subsequent software baseline correction, characteristic peak extraction and quantization processing.

[0009] Furthermore, the computer software processes the multi-band spectral data recorded by the portable multi-functional spectrometer in the following order: First, baseline correction and intensity normalization are performed on the ultraviolet-visible and near-infrared spectra; then, using the epidermal pigment absorption peak at 680 nm in the ultraviolet-visible reflectance spectrum as a background reference, adaptive filtering is performed on the fluorescence spectrum in the frequency or time domain to remove background components related to epidermal pigments; and polynomial baseline correction is performed on the Raman spectrum to subtract fluorescence baseline components; then, predetermined characteristic peak intensities are extracted from each of the above-processed spectral bands and constructed into feature vectors in a fixed order; the feature vectors are compared with a preset standard spectral library to identify the position and shape of candidate characteristic peaks, and the characteristic peak intensities confirmed by the comparison are used as input to a partial least squares regression (PLS) model or an artificial neural network (ANN) model, the model outputting the quantitative results of the target pesticide in the sample.

[0010] This processing flow clarifies the representation of the signal to be analyzed in different spectral bands through unified baseline correction, background subtraction based on 680nm pigment reference, and baseline processing of Raman spectra. It also inputs multi-spectral band features into the fusion quantitative model in a consistent format, forming a complete chain of spectral preprocessing-feature extraction-quantitative analysis.

[0011] This invention also provides an in-situ method for determining pesticide residue concentrations in fruits and vegetables. This method is based on the above-mentioned system and includes the following main steps: Sample preparation: Gently wipe the surface of the fruit to be tested with a soft brush or sterile cotton cloth to remove visible contaminants such as dirt and dust, without using chemical reagents or destructive treatment. Place the cleaned sample in the sample cell and fix it with the sample clamp, so that the surface to be tested is perpendicularly aligned with the optical axis of the multispectral fusion probe. When the sample to be tested is a liquid sample, pour it into a transparent cuvette and wipe the outer wall to eliminate the influence of outer wall reflection on the measurement. This is used to control the incident light path conditions and reduce the interference of surface contaminants on scattering and absorption of spectral acquisition. System calibration and SNR adjustment: Place pesticide standard samples of known concentrations into the system and start the system. Adjust the light source intensity and adjustable slit width, while the CMOS imaging module acquires signals and the computer software calculates the signal-to-noise ratio (SNR) in real time. Adjust the SNR to ≥30dB. The above calibration and threshold confirmation are used to establish the benchmark for acquisition conditions, which serve as a prerequisite for subsequent reliable feature extraction. Module selection and time-series acquisition: Based on the chemical characteristics of the target pesticide, the corresponding combination of spectroscopic modules is selected and acquisition is activated; the ultraviolet-visible module and the fluorescence module are acquired alternately in a 0.5-second time-division mode, and the fluorescence module is acquired with a 10ns time-resolved delay to distinguish transient emission from steady-state background; the near-infrared module and the enhanced Raman probe are acquired synchronously through fiber optic splitting. The raw spectra obtained by each module are recorded by a portable multi-functional spectrometer and transmitted to computer software, forming a multi-spectral parallel / time-division raw data stream according to a predetermined time sequence. The above time sequence arrangement is used to avoid excitation light crosstalk and ensure the time consistency of near-infrared and Raman data, which facilitates subsequent multi-spectral fusion processing; Preprocessing and background subtraction: The computer software sequentially performs baseline correction and intensity normalization on the original spectrum; using the epidermal pigment absorption peak at 680 nm in the UV-Vis reflectance spectrum as a background reference, adaptive filtering is applied to the fluorescence spectrum in the frequency or time domain to subtract the background component introduced by the epidermal pigment; polynomial baseline correction is applied to the Raman spectrum to remove the fluorescence background. This preprocessing sequence is used to unify the baselines and dimensions of different spectral bands and reduce the interference of epidermal pigment and fluorescence baselines on the identification of characteristic peaks; Feature extraction and matching: Predetermined characteristic peak intensities are extracted from processed UV-Vis, near-infrared, fluorescence, and Raman spectra, and the characteristic peak intensities are matched with fingerprint spectra in a preset standard spectral library to identify candidate target substances; the matched characteristic peaks are arranged in a fixed order to form a feature vector for use by the quantitative model; Fusion Quantitative Analysis: The above feature vectors are used as inputs, and fusion quantitative models such as partial least squares regression (PLS) or artificial neural networks (ANN) are used for calculation. The model outputs the type and concentration of each target pesticide in the sample, and the system generates a detection report. This fusion process uses multi-spectral complementary information as input to form a unified quantitative analysis path. Cleaning and data backup: After the test is completed, rinse the sample cell and probe surface with deionized water and wipe them dry. Back up the original spectral files and test reports for historical traceability and retesting.

[0012] Furthermore, the system calibration and SNR adjustment are performed according to the following procedure: a pesticide standard sample of known concentration is placed in the sample cell and the system is started. The CMOS imaging module acquires the spectrum of the standard sample. The computer software receives and processes the acquired signal to calculate the current signal-to-noise ratio (SNR). By adjusting the light source intensity and the adjustable slit width and re-measuring the spectral signal in real time, the adjustment process continues until the SNR calculated by the computer software reaches no less than 30dB. When the SNR reaches this value, the module selection and timing acquisition are entered according to the predetermined procedure. This calibration establishes a quantitative criterion for the acquisition conditions based on the standard sample, and uses this criterion as the trigger condition for subsequent multi-spectral acquisition, thereby quantitatively confirming the incident light path and signal quality before acquisition.

[0013] Furthermore, the module selection and timing acquisition are configured according to the target pesticide category. For organophosphorus pesticides, a combination of ultraviolet-visible spectrophotometry module, near-infrared spectroscopy module, and enhanced Raman probe is used for acquisition; for pyrethroid pesticides, a combination of fluorescence spectroscopy excitation module, near-infrared spectroscopy module, and enhanced Raman probe is used for acquisition. The UV-Vis spectrophotometer module and the fluorescence spectroscopy excitation module are acquired alternately at 0.5-second intervals to prevent crosstalk between excitation sources. The near-infrared spectroscopy module and the enhanced Raman probe are synchronously acquired via fiber optic splitting to ensure that near-infrared vibrational absorption information and Raman fingerprints are acquired and stored in parallel within the same time window. The above configuration is based on the complementarity of each spectral band for different chemical characteristics: the UV-Vis channel is used to record electronic absorption features, the fluorescence channel is used to record fluorescent targets, the near-infrared channel is used to record molecular vibrational bands, and the Raman channel is used to record molecular fingerprints. The multi-band raw data acquired according to the above timing and synchronization strategy are input in parallel in a unified format in the subsequent software processing stage for baseline correction, feature extraction, and multi-band fusion quantitative analysis.

[0014] Further, the preprocessing and background subtraction are performed in the following order: First, baseline correction and intensity normalization are performed on the UV-Vis and near-infrared spectra; then, the epidermal pigment absorption peak at 680 nm is identified and located in the UV-Vis reflectance spectrum, and a background curve is established in the neighborhood of this peak through local polynomial fitting or adaptive filtering; the background curve is interpolated according to the wavelength correspondence and the corresponding background component is subtracted from the fluorescence spectrum point by point; polynomial baseline correction is performed on the Raman spectrum after background subtraction to eliminate residual fluorescence baseline; after completing the above processing, the intensity of characteristic peaks is extracted from the processed UV-Vis, near-infrared, fluorescence and Raman spectra according to predetermined rules and a feature vector is constructed. This processing sequence forms a consistent and comparable multi-band feature input by first unifying the spectral baseline and dimensions, then fitting based on the 680 nm pigment peak and subtracting the background in the fluorescence band, and finally performing baseline correction on the Raman spectrum, for subsequent feature matching and fusion quantitative model use.

[0015] Furthermore, the fusion quantification is implemented as follows: the intensity of multi-spectral characteristic peaks obtained from preprocessing and feature extraction is arranged in a fixed order to form a feature vector, which serves as the input to the quantitative model; the quantitative model can be implemented using partial least squares regression (PLS) or artificial neural network (ANN), where PLS, as a linear multivariate regression method, processes the multicollinearity of spectral data by projecting to a low-dimensional latent variable space, and is suitable for modeling the approximate linear relationship between features and concentrations; ANN is a feedforward network that can express nonlinear mappings, and in this invention, the ANN is exemplarily composed of an input layer, three hidden layers, and an output layer, with each output node of the output layer corresponding to the concentration value of each target pesticide in the sample to be tested.

[0016] The beneficial effects of this invention are as follows: 1. This invention employs a modular multi-spectral probe and controllable timing in this system. The ultraviolet-visible spectrum is used as a baseline reference for epidermal pigments, the fluorescence channel is time-resolved for acquisition, and near-infrared and enhanced Raman are acquired synchronously through optical coupling. With the positioning of the sample holder and the non-destructive optical path arrangement, stable and repeatable spectral sampling of the fruit surface can be performed in situ. This helps to reduce the interference of matrix pigments and background fluorescence on the target signal and reduce measurement fluctuations caused by sample non-uniformity.

[0017] 2. This invention performs baseline correction, intensity normalization, and background subtraction based on skin reference on each spectral band sequentially. After extracting characteristic peaks, a feature vector is constructed according to a fixed index. Quantitative calculation is then completed using a linear or nonlinear fusion model. The above processing flow can comprehensively utilize complementary information from multiple spectral bands, improve the stability of target object identification and quantification under complex matrices, reduce erroneous judgments caused by matrix interference, and enhance the consistency and reproducibility of on-site quantitative results.

[0018] 3. This invention provides a portable, modular system with replaceable and maintainable key optical components and enhancement substrates. At the software level, calibration procedures and acquisition quality thresholds are set, and acquisition parameters and model versions are recorded. This configuration facilitates rapid on-site screening and maintains the traceability of the testing process and results. It provides a verifiable data chain for laboratory retesting and regulatory judgment when necessary, reducing reliance on centralized laboratories. Attached Figure Description

[0019] Figure 1 This is a flowchart of the core detection process of the system of the present invention; Figure 2 This is a flowchart of the spectral preprocessing and background subtraction process of the present invention; Figure 3 This is a flowchart of the module selection and timing acquisition process for this invention; Figure 4 This is the spectrum of a 100% concentration potassium sulfide solution according to the present invention; Figure 5 This is a spectral image of the sample from the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figures 1 to 5 As shown, this embodiment of the invention provides an in-situ determination system for pesticide residue concentration in fruits and vegetables. The detection process is completed in a portable multifunctional spectroscopic testing system based on spectral analysis. The system includes: The system includes: a sample cell, a sample holder, a light source, a multispectral fusion probe, a portable multifunctional spectrometer, and computer software for communicating with the spectrometer. The multispectral fusion probe has a modular structure and includes: a UV-Vis spectrophotometric module (operating wavelength 200–760 nm), a near-infrared spectroscopy module (operating wavelength 900–2500 nm), a fluorescence spectroscopy excitation module (excitation wavelength 280–450 nm), and an enhanced Raman probe (Raman shift scanning range 200–2000 cm⁻¹). -1 The enhanced Raman probe incorporates a SERS substrate primarily composed of gold nanoparticles with a particle size of approximately 50 nm. These gold nanoparticles are electrostatically self-assembled and immobilized on the surface of a silicon wafer carrier. The SERS probe is connected to a portable spectrometer via optical fiber or an optical coupler, and the Raman signal is transmitted to the spectral recording unit via optical fiber. Optical Recording Unit and Key Parameters: The optical recording chain of the portable multi-functional spectrometer includes: an adjustable slit, a biconvex lens (focal length approximately 3 cm), a grating (specification example 20 mm × 20 mm × 2 mm, line density example 1000 lines / mm), and a high-resolution CMOS imaging module (effective pixels example 5 million, pixel size example 2.2 μm). The slit width is adjusted by a micrometer head driving a pair of blades to make a small linear motion (initial slit width example 50 μm). The above optical components are registered according to the matching relationship between spectral resolution and pixel sampling during assembly to ensure that the spatial distribution of wavelength on the CMOS can be resolved and quantitatively processed. SERS substrate fabrication and probe assembly: Silicon wafer pretreatment: The optical grade silicon wafers were cleaned with acetone, ethanol and deionized water respectively, for 5 minutes each; dried under clean conditions and stored under nitrogen protection; Gold nanoparticle treatment: Gold nanoparticles with a particle size of 50 nm were used as a storage solution; before use, the solution was diluted with deionized water and ultrasonically dispersed for 5 min. Electrostatic self-assembly: The treated silicon wafer is immersed in a gold nanoparticle suspension and left to stand to complete self-assembly. Then it is gently rinsed with deionized water and dried at low temperature to obtain a uniformly distributed gold nanoparticle layer as the SERS substrate. Probe assembly: The prepared silicon wafer substrate is fixed to the base of the enhanced Raman probe, the optical coupling position is adjusted and the particle coverage quality is confirmed by microscopic inspection, and the fiber coupling and path verification are completed. Maintenance Instructions: The SERS substrate is a replaceable part. Replacement should be performed and recorded according to the maintenance plan (the gold nanoparticle substrate should be replaced monthly to ensure SERS activity; the replacement batch and date should be recorded in the maintenance log). Sample preparation and system calibration: Sample preparation: Use a soft brush or sterile cotton cloth to gently wipe the surface of the fruit to be tested to remove visible contaminants, without using chemical reagents or destructive treatments; Place the cleaned fruit in the sample cell and fix it with the sample clamp, ensuring that the surface to be tested is perpendicular to the optical axis of the probe; When testing liquid samples, inject the sample into a transparent cuvette and wipe the outer wall to eliminate the influence of air bubbles and reflection from the outer wall; System calibration and SNR adjustment: Place pesticide standard samples of known concentration in the sample cell, start the system and acquire data according to the initial settings (slit 50μm, UV light source intensity 50–100mW); the spectrum is recorded by the CMOS imaging module and the SNR is calculated by computer software; by adjusting the light source intensity and slit width and repeating the acquisition until the SNR ≥ 30dB, the formal acquisition program is entered. The SNR can be calculated by taking the peak signal S at the target peak position and the standard deviation σ of the noise in the adjacent baseline area, and calculating SNR(dB) = 20·log10(S / σ); Module selection and data acquisition timing: Select the module combination based on the type of pesticide to be tested and enable data acquisition: Organophosphorus pesticides: Utilize a combination of ultraviolet-visible module, near-infrared module, and enhanced Raman probe; Pyrethroid pesticides: A combination of fluorescence module, near-infrared module, and enhanced Raman probe is used. The UV-Vis and fluorescence modules are acquired alternately in a time-division manner with a time interval of 0.5s. The fluorescence module performs time-resolved acquisition with a 10ns delay after excitation to reduce background influence. The near-infrared module and the enhanced Raman probe are acquired synchronously in the same time window through an optical fiber splitter. The raw spectra acquired by each module are recorded by the spectral recording unit and transmitted to the computer software in real time for subsequent processing. Spectral preprocessing and background subtraction Baseline correction and normalization: Baseline correction is performed on the ultraviolet-visible and near-infrared spectra, and intensity is normalized according to the peak value or spectral integral; Epidermal pigment reference and fluorescence background subtraction: Identify the epidermal pigment absorption peak at 680 nm in the UV-Vis reflectance spectrum and establish a background curve in this neighborhood by local fitting or adaptive filtering. Interpolate the background curve according to the wavelength correspondence and subtract the corresponding components from the fluorescence spectrum point by point. Raman spectrum baseline correction: Polynomial baseline correction is performed on the Raman spectrum to eliminate fluorescence baseline residue and can be supplemented with smoothing filtering; Feature extraction: Locate and record the peak position, peak height and peak area of ​​the characteristic peaks in each spectral band according to predetermined rules or derivative method, and construct feature vectors in a fixed order for model input; Model training and validation: The characteristic peak intensities of each band obtained through feature extraction and matching are used to form a feature vector in a fixed order, and this feature vector is used as the input of a partial least squares regression (PLS) model or an artificial neural network (ANN) model to perform multi-band fusion quantification. Data preprocessing and set partitioning: Baseline correction, peak normalization or spectral integral normalization, and necessary noise smoothing are uniformly performed on the feature vectors obtained from preprocessing; then, the available samples are divided into training set, validation set and test set according to the ratio (for example, the ratio is 70%:15%:15%, which can be adjusted according to the sample size). All training inputs are standardized according to a uniform scale before entering the model to avoid the influence of differences in units on model training. PLS Modeling Process (Linear Baseline) Perform partial least squares regression (PLS) on the training set; determine the number of latent variables through cross-validation (e.g., 5-fold cross-validation is used in the example), and use the root mean square error (RMSE) on the validation set as the main criterion. After determining the number of latent variables, train the final model on the training set and evaluate its coefficient of determination R² and RMSE on the test set as a benchmark evaluation of the model's linear performance. ANN Modeling Process (Nonlinear Options) A feedforward artificial neural network (ANN) is constructed as a nonlinear modeling scheme. For demonstration purposes, the ANN in this embodiment consists of an input layer, three hidden layers, and an output layer. Each output node in the output layer corresponds to the concentration of the target pesticide. The demonstrative training strategy is as follows: the number of nodes in the input layer is equal to the dimension of the feature vector; the demonstrative number of hidden layer nodes is 128, 64, and 32; the loss function is the mean squared error (MSE); the Adam optimizer is used for training, with an initial learning rate of 1×10⁻³, a batch size of 32, a maximum number of training epochs of 200, and an early stopping strategy to prevent overfitting; the hyperparameters are determined by grid search or Bayesian optimization on the validation set, and the model stability is evaluated by cross-validation. After training, the coefficient of determination R², RMSE, and residual distribution are calculated on the test set to evaluate the generalization performance.

[0022] Among them, module definition and working scope The multispectral fusion probe contains the following spectroscopic modules and operates within the following working / scanning ranges: UV-Vis spectrophotometer module: working wavelength 200–760 nm; used to record the electronic absorption characteristics in the reflectance spectrum of the sample surface and identify epidermal pigment absorption peaks; Near-infrared spectroscopy module: operating wavelength 900–2500 nm; used to record molecular vibrational absorption bands (such as C–H, O–H vibrational information). Fluorescence spectroscopy excitation module: excitation wavelength 280-450nm; used to excite and collect the emission spectrum of target objects with fluorescence response, and combined with time-resolved acquisition to suppress epidermal fluorescence background; Enhanced Raman probe: Raman shift scanning range 200–2000 cm⁻¹ -1 A SERS substrate was used to obtain a magnified Raman fingerprint spectrum. SERS Substrate and Coupling Method: The enhanced Raman probe incorporates a gold nanoparticle substrate (50 nm in diameter) that is electrostatically self-assembled and fixed to the surface of a silicon wafer. This substrate is assembled inside the Raman probe as a replaceable component. The Raman scattering signal is transmitted to the portable spectral recording unit through an optical fiber or optical coupler. The preparation, assembly, and replacement procedures of the substrate are detailed in the specific implementation. Acquisition timing and synchronization control (for coordinated acquisition of different spectral bands): The system adopts the following timing and synchronization strategy: the UV-Vis module and the fluorescence module acquire data alternately in a time-division manner, with an alternation interval of 0.5 seconds; the fluorescence module performs time-resolved acquisition, with a 10ns delay after excitation before acquisition; the near-infrared module and the enhanced Raman probe acquire data synchronously within the same time window through optical coupling (fiber optic splitting or equivalent coupling). The above timing ensures the correspondence of data in each spectral band in physical sampling, facilitating subsequent fusion analysis; System calibration is performed before acquisition: pre-acquisition is performed using standard samples of known concentrations, and the spectrum is recorded by the CMOS imaging module. The signal-to-noise ratio (SNR) is calculated by computer software. The SNR is determined by the ratio of amplitude to baseline noise standard deviation and expressed in dB; formal acquisition only begins when the SNR reaches or exceeds 30dB. During calibration, the SNR requirement can be met by adjusting the slit width (driven by a micrometer screw, initial slit width 50μm) and the light source intensity (UV light source intensity range 50–100mW), and the calibration settings and SNR values ​​are recorded for traceability. Preprocessing and background subtraction (explicit steps within the working range of each spectral band) involve processing the raw spectra of each spectral band and extracting features in the following order: Baseline correction and intensity normalization were performed on the UV-Vis and near-infrared spectra; the epidermal pigment absorption peak at approximately 680 nm was identified and located in the UV-Vis spectrum, and a background reference curve was generated accordingly. After time-resolved acquisition of fluorescence spectra, the corresponding background components are subtracted point by point based on the background reference curve constructed by ultraviolet-visible using an adaptive filtering method. Polynomial baseline correction was performed on the Raman spectrum to eliminate fluorescence baseline residue.

[0023] The enhanced Raman probe incorporates a surface-enhanced Raman scattering (SERS) substrate. The SERS substrate is composed of gold nanoparticles with a particle size of 50 nm. The gold nanoparticles are electrostatically self-assembled onto the surface of a silicon wafer to form a continuous or near-continuous nanoparticle coating layer. This coating layer serves as the active surface of the SERS, amplifying the Raman scattering signal of trace pesticide molecules on the sample surface (with an enhancement factor of up to 10^6 in some examples). The SERS substrate is mounted in the base of the enhanced Raman probe, with the active surface of the probe facing the sample surface and maintaining replaceability. In terms of signal transmission, the output end of the enhanced Raman probe is connected to the incident port of the portable multifunctional spectrometer via optical fiber or equivalent optical coupler. The Raman scattered light collected by the probe is transmitted to the portable spectrometer recording unit for reception and recording via the coupler to ensure the comparability and stability of the Raman data in subsequent processing.

[0024] The optical recording chain of the portable multifunctional spectrometer, in the order of optical path, includes: an adjustable slit, a lens, a grating, and a high-resolution CMOS imaging module. Its specific structure and implementation are as follows: Device Specifications Adjustable slit: Consists of a micrometer screw, a spring, and a pair of blades; the micrometer screw drives the blades to move in a straight line to change the slit width; the initial slit width is set to 50 μm; Lenses: Two biconvex lenses, each with a focal length of 3cm, used for collimation and imaging; Grating: Sinusoidal original grating, size 20mm×20mm×2mm, line density 1000 lines / mm, used for dispersive incident light to achieve wavelength separation; CMOS imaging module: approximately 5,000,000 effective pixels, pixel size approximately 2.2μm, used to record the dispersed spectral bands and transmit them to computer software for digital processing; Optical Assembly and Positioning: Each optical element is fixed coaxially within the optical cavity or modular housing. Incident light enters through a slit, is collimated by a first biconvex lens, and then incident on the grating. The grating disperses the spectrum, which is then converged by a second lens. The dispersed spectral bands are imaged on the CMOS array. The relative positions of the slit, lens, grating, and CMOS are fixed during assembly using mechanical positioning references and precisely registered using an optical axis alignment procedure during factory testing or maintenance to ensure a stable positional relationship of the spectral lines on the CMOS pixels. Slit Adjustment and Signal-to-Noise Ratio Considerations: The slit width is adjusted by a micrometer screw. Rotating the micrometer screw causes the blade to move linearly, thus changing the opening width. The slit is set to an initial width of 50μm as the default acquisition condition during factory testing or commissioning. In the field or during calibration, the slit can be fine-tuned as needed to achieve a balance between spectral resolution and signal intensity. In the acquisition process of this system, the signal-to-noise ratio (SNR) calculated in real time by the computer software should be used as the basis. Once the SNR reaches the specified threshold, the acquisition parameters are determined. Focusing and image plane calibration process: Preliminary alignment: After installing the optical components, use the coarse adjustment mechanism to align the center of the grating with the center of the CMOS; Fine focusing: Adjust the axial position of the lens and the CMOS to make the dispersed spectral bands appear as clear and continuous stripes on the CMOS; Pixel-wavelength mapping: Use a reference signal or standard sample with a known wavelength to calibrate the pixel to the wavelength, and record the mapping coefficients for automatic conversion by the software during daily acquisition. Readout and Data Interface: The digital output of the CMOS imaging module is transmitted to the signal processing unit in the portable multi-functional spectrometer or an external computer via USB or equivalent data interface. The signal processing unit performs bad pixel correction, row and column correction and amplitude quantization and stores the raw spectral data as a standard data file for subsequent processing. Cleaning, maintenance, and replacement guidelines: To maintain optical performance and measurement repeatability, inspect and clean the slit, lens, and grating surfaces monthly. Use solvents and methods safe for optical components when cleaning (e.g., rinse with deionized water first, then wipe gently with a lint-free cloth). If scratches or contamination on the grating or lens cannot be repaired, replace them according to the replacement procedure and recalibrate the image plane. The mechanical transmission components of the slit (micrometer head, spring, blade) should be checked for wear and return accuracy regularly, and maintenance records should be included in the equipment maintenance log.

[0025] The raw ultraviolet-visible, near-infrared, fluorescence, and Raman spectral data recorded by the portable multi-functional spectrometer are processed by computer software in the following order, and eigenvectors are output for quantitative analysis: Raw data reception and time alignment The original spectral files of each spectral band are received and stored according to the acquisition timestamp and channel number. Each record also saves the acquisition parameters (slit width, light source intensity, acquisition time sequence identifier, SNR value, etc.) as a basis for subsequent processing and traceability. Baseline correction and intensity normalization (UV-Vis, NIR): The UV-Vis and NIR spectra are first baseline corrected, and polynomial fitting or spline fitting is used to remove baseline drift. Then, the amplitude of the corrected spectrum is normalized. The normalization method is peak normalization or spectral integral normalization to eliminate amplitude differences under different acquisition conditions. Epidermal pigment reference and fluorescence background subtraction: The epidermal pigment absorption peak at approximately 680 nm is identified and located in the corrected UV-Vis spectrum. A background reference curve is constructed based on this neighborhood. The background reference curve is mapped to the fluorescence band and the corresponding background component of the fluorescence spectrum is subtracted point by point using adaptive filtering to complete fluorescence background suppression. Raman spectrum baseline correction and defluorination: Polynomial baseline correction was performed on the acquired Raman spectrum to eliminate residual fluorescence baseline. If necessary, smoothing filtering was used to process high-frequency noise. After correction, baseline normalization was performed for comparison and matching. Peak location and feature extraction: The derivative method and local extremum method are used to locate the characteristic peaks of each spectral segment; the peak position, peak height and peak area of ​​each characteristic peak are recorded as feature quantities, and the characteristic peaks of each spectral segment are arranged in a predetermined index order and merged into a unified feature vector; Fingerprint spectral matching: The extracted feature peaks are matched with the standard fingerprint spectral library. The matching is based on the consistency of peak position and relative intensity distribution. The matching threshold and tolerance are executed according to the calibration parameters. The intensity of the feature peaks confirmed by the matching is retained as valid input. Peaks that do not meet the matching threshold are marked according to the rules and enter the review or subsequent processing. Quantitative input and concentration calculation: The matched feature vectors are input into the partial least squares regression (PLS) model or artificial neural network (ANN) model for quantitative calculation. The model output is the concentration value of each target pesticide (unit: ppb). The model file contains the model version number and training records. The quantitative calculation outputs the model version number and confidence index at the same time. Results Output and Archiving: Generate a test report, which includes sample identification, acquisition settings, processing parameters, model version number, comparison conclusions of each target pesticide concentration with the standard, and archives the original spectrum, processed data, feature vectors, model inputs and outputs along with the test report for easy retesting and traceability.

[0026] This invention also provides an in-situ method for determining pesticide residue concentrations in fruits and vegetables. This method is based on the above-mentioned system and includes the following main steps: Sample preparation: Use a soft brush or sterile cotton cloth to gently wipe the surface of the fruit to be tested to remove visible contaminants without using chemical reagents or destructive treatments. Place the cleaned fruit in the sample cell and fix it with the sample clamp so that the surface to be tested is perpendicularly aligned with the optical axis of the multispectral fusion probe. When the sample to be tested is a liquid sample, pour it into a transparent cuvette and wipe the outer wall. System calibration and SNR adjustment: Place pesticide standard samples of known concentration into the sample cell and start the system. Adjust the light source intensity and adjustable slit width, and use computer software to monitor the SNR of the CMOS acquisition signal in real time to adjust the SNR to ≥30dB. Module selection and timing acquisition: Select the combination of spectroscopic modules according to the type of target pesticide and enable acquisition. The ultraviolet-visible module and the fluorescence module are acquired alternately in a 0.5-second time-division mode. The fluorescence module is acquired with a 10-ns delay and time resolution. The near-infrared module and the enhanced Raman probe are acquired synchronously through fiber optic splitting. The original spectra obtained by each module are recorded by a portable multi-functional spectrometer and transmitted to the computer software. Preprocessing and background subtraction: The computer software performs baseline correction and normalization on the original spectrum in sequence, and performs adaptive filtering with the epidermal pigment absorption peak at about 680 nm as a reference to subtract the background. The Raman data is subjected to polynomial baseline correction to eliminate the fluorescence background. Feature extraction and matching: Extract the intensity of characteristic peaks from the processed spectra of each band and match them with a preset standard spectral library; Fusion Quantification: The intensity of the characteristic peaks obtained from feature extraction and matching is input into the PLS or ANN fusion model for quantitative analysis, outputting the type and concentration of each target pesticide in the sample and generating a detection report; Cleaning and data backup: After the test is completed, rinse the sample cell and probe surface with deionized water and wipe them dry. Back up the original spectral files and test reports for historical traceability and retesting.

[0027] In the system calibration and SNR adjustment, pesticide standard samples of known concentrations were used for system calibration. The standard samples were placed in the sample cell and fixed with sample clamps to ensure that the sample surface was perpendicularly aligned with the optical axis of the multispectral fusion probe. Liquid standard samples were injected into transparent cuvettes and the outer walls were wiped until no visible stains were visible. Acquisition preparation and initial settings: The portable multi-functional spectrometer and CMOS imaging module were started, the initial slit width was set to 50 μm, the light source was started according to the equipment's nominal value (UV light source intensity setting range 50–100 mW), and the initial acquisition parameters were recorded. The SNR calculation method: The target spectral band signal of the standard sample was acquired by the CMOS imaging module. The software selected the characteristic peak region and the adjacent baseline region in the target spectral band, calculated the signal amplitude S and the baseline noise standard deviation σ, and calculated the signal-to-noise ratio according to the following formula: SNR(dB)=20·log10(S / σ); the signal S is the peak intensity or peak area of ​​the characteristic peak, and the noise σ is the standard deviation of the intensity of the baseline area. The software records the values ​​of S, σ and SNR in a preset manner. Adjustment process (to achieve closed-loop control with SNR ≥ 30dB): The software reads the current SNR; if SNR < 30dB, adjust and re-acquire in the following order: increase the light source intensity (adjust within the equipment's allowable range); moderately increase the slit width (achieve this through fine-tuning the blade of the micrometer screw); re-acquire and calculate SNR after each adjustment until SNR ≥ 30dB; when SNR ≥ 30dB, use the current acquired parameters as the valid calibration parameters for this test and record them, then start the module selection and timing acquisition process; Handling and recording calibration failures: If the SNR still cannot reach 30dB after multiple adjustments in the above order, record the current acquisition parameters, SNR and environmental status information, stop entering the formal acquisition process and generate a calibration failure prompt for maintenance or recalibration. All calibration records (including standard sample concentration, slit width, light source intensity, acquisition timestamp, SNR value and final judgment) are stored in the system log for traceability.

[0028] The module selection and timing acquisition are configured and implemented according to the pesticide category to be tested. Before acquisition, the system calibration should be completed and the acquisition prerequisites such as SNR≥30dB and initial slit width 50μm should be confirmed. Module combination rules: Organophosphorus pesticides: Use the UV-Vis spectrophotometry module, near-infrared spectroscopy module, and enhanced Raman (SERS) probe simultaneously for detection; Pyrethroid pesticides: Use the fluorescence spectroscopy excitation module, near-infrared spectroscopy module, and enhanced Raman (SERS) probe simultaneously for detection; Timing control and synchronization: The UV-Vis and fluorescence modules are acquired alternately in a time-division manner, with an alternation interval of 0.5 seconds; at any given time, only the UV-Vis or fluorescence module is in the excitation / recording state to avoid excitation light crosstalk; the fluorescence module uses time-resolved acquisition; the emission signal is acquired after a 10ns delay after excitation to reduce the influence of short-lived background components on the target fluorescence peak; the near-infrared module and the enhanced Raman probe are synchronously acquired within the same time window through fiber optic splitting or equivalent optical coupling, and the data from both correspond to the same physical sampling time, which facilitates subsequent multi-spectral fusion processing; Acquisition Sequence Diagram (Single Detection Cycle): System Start: Confirm current acquisition parameters (slit width, light source intensity, etc.) and SNR ≥ 30dB; If the target is organophosphates, acquire data in the following parallel / serial manner: Synchronous acquisition: The near-infrared module and SERS probe acquire data synchronously within time window T and save the raw data; Time-division acquisition: The UV-Vis module alternates with the fluorescence module at 0.5s intervals before and after the near-infrared / SERS synchronization window (the fluorescence module remains unexcited when not in use); If the target is pyrethroids, acquire data in the following parallel / serial manner: Synchronous acquisition: The near-infrared module and SERS probe acquire data synchronously within time window T and save the raw data; Time-division acquisition: The fluorescence module and UV-Vis module alternate at 0.5s intervals before and after the near-infrared / SERS synchronization window (the fluorescence module remains unexcited when not in use); Time-alternating acquisition (10ns delay acquisition after fluorescence module excitation); Acquisition parameters and records: During the acquisition process, the timestamp, module status (enabled / disabled), light source intensity, slit width, SNR value, and synchronization identifier of each acquisition are recorded and archived. The synchronous acquisition of near-infrared and SERS data files are marked with the same time window for subsequent fusion analysis; Post-processing interface: The acquired multi-band raw spectra are integrated according to the time window and transferred to the computer software; In the software, the epidermal pigment absorption peak at approximately 680nm in the UV-Vis spectrum is used as the background reference. The fluorescence and Raman data are subjected to corresponding background subtraction and baseline correction according to the established preprocessing procedure. Then, the characteristic peaks are extracted and input into the fusion quantitative model (PLS or ANN) for quantitative calculation.

[0029] Among them, the original multi-band spectra are preprocessed and background subtracted in the following order, and the characteristic peak intensities are extracted from each band. Baseline correction and intensity normalization of UV-Vis and near-infrared spectra: Baseline correction and intensity normalization were performed on UV-Vis and near-infrared spectra respectively. Baseline correction adopted a polynomial fitting method: the baseline of the spectrum to be measured was fitted with a second-order polynomial across the entire wavelength range and the fitted baseline was subtracted from the original spectrum. After baseline subtraction, the spectral amplitude was normalized according to the peak value to make the normalized spectral intensity range uniform for subsequent comparison and fusion. The processing results were output as preprocessed data of UV-Vis and near-infrared spectra and archived. A background reference was constructed based on the 680 nm pigment peak, and adaptive filtering was applied to the fluorescence spectrum. The epidermal pigment absorption peak at approximately 680 nm was located in the baseline-corrected UV-Vis reflectance spectrum. The background reference curve was generated by local fitting using the absorption peak and its neighborhood data. After interpolating and mapping the background reference curve to the wavelength coordinates of the fluorescence emission band, the corresponding background component was subtracted from the fluorescence spectrum point by point using adaptive filtering to complete the fluorescence background subtraction. The fluorescence spectrum background subtraction data was then normalized and archived. Baseline correction and defluorination of Raman spectra: Polynomial baseline correction is first performed on the Raman spectrum: the baseline of the Raman spectrum is fitted with a third-order polynomial and the obtained baseline is subtracted to eliminate the fluorescence baseline component; after baseline subtraction, the Raman spectrum is smoothed and filtered to suppress high-frequency noise. The Raman spectrum after baseline correction and filtering is output as Raman preprocessing data and archived. Location and intensity extraction of characteristic peaks in each spectral band: In the preprocessed data, characteristic peaks are located in the ultraviolet-visible, near-infrared, fluorescence and Raman spectra respectively using the derivative method and the local extremum method; for each located characteristic peak, three characteristic quantities are recorded: peak position (wavelength or Raman shift), peak height and peak area. The recorded characteristic quantities of each spectral band are merged into a unified feature vector in a predetermined order. Results Recording and Traceability: All preprocessing parameters (baseline polynomial order, neighborhood range used to construct epidermal pigment background, mapping interpolation method, filtering method and parameters, peak location method and threshold, etc.) are recorded in the detection log along with the processing results; the original spectrum, processed spectrum, feature vector and processing log are archived for re-examination. The above records ensure the reproducibility and traceability of this implementation method.

[0030] The process of feature vector-fusion model-concentration output is used to achieve multi-spectral quantification, as follows: Input data: The feature vector obtained by the preprocessing and background subtraction steps is used as the model input. The feature vector is composed of the characteristic peak intensities of each spectral band (UV-Vis, near-infrared, fluorescence, Raman) in a fixed index order. Before input, the feature vector is uniformly scaled (zero mean or variance standardization). Model selection: Quantitative calculations are performed using partial least squares regression (PLS) or artificial neural networks (ANN). Either one can be used for field inference if it meets the acceptance criteria. If both meet the criteria, the preferred deployment scheme is determined based on the test set RMSE and computational efficiency. PLS modeling: Perform PLS on the training set: determine the number of latent variables through cross-validation and use RMSE as the criterion on the validation set; calculate the coefficient of determination R² and RMSE on the test set, and record the model parameters and training logs; ANN Modeling: A feedforward ANN is constructed, consisting of an input layer (number of nodes equal to the feature vector dimension), three hidden layers, and an output layer. Each output node in the output layer corresponds to the concentration (unit: ppb) of the target pesticide. Hyperparameters are determined through cross-validation. During training, early stopping is performed based on validation set metrics to select the final model. R², RMSE, and the limit of detection (LOD) are recorded on the test set.

[0031] Acceptance and Version Management: The model acceptance threshold is: the coefficient of determination R² on the test set or independent validation set is ≥0.99, and the detection limit LOD is ≤1ppb. Models that pass the acceptance are saved in the form of version numbers. The model file contains the model structure, training timestamp, training log and performance metrics. Model replacement is only performed after the acceptance threshold is met. On-site inference and recording: When the model is deployed for on-site inference, the system receives the processed feature vectors and outputs the concentration values ​​of each target pesticide; the inference results, the model version number used, the corresponding original spectrum and the processed data are saved together, and the test report lists the sample identification, the concentration of each target pesticide and the comparison conclusion with the applicable standard; Retraining Triggering Conditions and Traceability: Retraining is triggered when any of the following occurs: a large number of new calibration samples are added, the optical path or probe is replaced, resulting in a change in the system response, the SERS substrate is replaced, or a systematic deviation occurs during continuous on-site detection. The model can only be replaced and put online after retraining is completed and the acceptance threshold is met. Performance reports are generated and archived for each training, verification, and online deployment to ensure traceability.

[0032] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An in-situ determination system for pesticide residue concentration in fruits and vegetables, characterized in that: The detection process is completed in a portable, multifunctional spectroscopic testing system based on spectral analysis, the system comprising: A sample cell for holding the fruit or liquid sample to be tested; a sample clamp for ensuring stable placement of the sample and positioning the test area; a light source; a multispectral fusion probe; a portable multi-functional spectrometer; and computer software communicating with the portable multi-functional spectrometer. The multispectral fusion probe has a modular structure, comprising a UV-Vis spectrophotometric module, a near-infrared spectroscopy module, a fluorescence spectroscopy excitation module, and an enhanced Raman probe. The reflectance spectrum acquired by the UV-Vis spectrophotometric module is recorded by the spectral recording unit and used by the computer software as an epidermal pigment baseline reference for background subtraction of the signals acquired by the fluorescence module and the enhanced Raman probe. The near-infrared spectroscopy module and the enhanced Raman probe are synchronously acquired through optical coupling. The raw spectra recorded by the spectral recording unit of each module are transmitted to the computer software as input for fusion quantitative analysis.

2. The in-situ determination system and method for pesticide residue concentration in fruits and vegetables according to claim 1, characterized in that: The working wavelength range of the ultraviolet-visible spectrophotometer module is 200-760 nm; the working wavelength range of the near-infrared spectrum module is 900-2500 nm; the excitation wavelength range of the fluorescence spectrum excitation module is 280-450 nm; and the Raman displacement scanning range of the enhanced Raman probe is 200-2000 cm -1 .

3. The in-situ determination system and method for pesticide residue concentration in fruits and vegetables according to claim 2, characterized in that: The enhanced Raman probe incorporates a gold nanoparticle substrate with a particle size of 50 nm. The gold nanoparticles are immobilized on the silicon wafer surface via electrostatic self-assembly. The Raman signal acquired by the enhanced Raman probe is transmitted to the incident port of the portable multifunctional spectrometer via optical fiber.

4. The in-situ determination system and method for pesticide residue concentration in fruits and vegetables according to claim 3, characterized in that: The portable multifunctional spectrometer sequentially includes an adjustable slit, a lens, a grating, and a high-resolution CMOS imaging module. The adjustable slit includes a micrometer screw, a spring, and a pair of blades. The micrometer screw drives the blades to move in a straight line to change the slit width, with the initial slit width set to 50 μm. The lens consists of two biconvex lenses with a focal length of 3 cm. The grating measures 20 mm × 20 mm × 2 mm and has a line density of 1000 lines / mm. The high-resolution CMOS imaging module has an effective pixel count of 5 million pixels and a pixel size of 2.2 μm.

5. The in-situ determination system for pesticide residue concentration in fruits and vegetables according to claim 4, characterized in that: The computer software sequentially performs the following processing steps on the multi-band spectral data recorded by the portable multifunctional spectrometer: baseline correction and intensity normalization are performed on the ultraviolet-visible and near-infrared spectra; a background reference is constructed based on the epidermal pigment absorption peak at 680 nm in the ultraviolet-visible reflectance spectrum, and adaptive filtering is performed on the fluorescence spectral data; polynomial baseline correction is performed on the Raman spectral data; characteristic peak intensities are extracted from the processed spectrum of each band, and the characteristic peak intensities are matched with fingerprint spectra in a preset standard spectral library. The matched characteristic peak intensities are used as input to a partial least squares regression model or an artificial neural network model to calculate the concentration of pesticide residues in the sample.

6. A method for in-situ determination of pesticide residue concentration in fruits and vegetables, characterized in that: The method is based on the system of claim 5, and the method includes: Sample preparation: Gently wipe the surface of the fruit to be tested with a soft brush or sterile cotton cloth to remove visible contaminants without using chemical reagents or destructive treatments. Place the cleaned fruit in the sample cell and fix it with the sample clamp so that the surface to be tested is perpendicularly aligned with the optical axis of the multispectral fusion probe. When the sample to be tested is a liquid sample, pour it into a transparent cuvette and wipe the outer wall. System calibration and SNR adjustment: Place pesticide standard samples of known concentration into the sample cell and start the system. Adjust the light source intensity and adjustable slit width, and use computer software to monitor the SNR of the CMOS acquisition signal in real time to adjust the SNR to ≥30dB. Module selection and timing acquisition: Select the combination of spectroscopic modules according to the type of target pesticide and enable acquisition. The ultraviolet-visible module and the fluorescence module are acquired alternately in a 0.5-second time-division mode. The fluorescence module is acquired with a 10ns delay and time resolution. The near-infrared module and the enhanced Raman probe are acquired synchronously through an optical fiber splitter. The original spectra obtained by each module are recorded by a portable multi-functional spectrometer and transmitted to the computer software. Preprocessing and background subtraction: The computer software performs baseline correction and normalization on the original spectrum in sequence, and performs adaptive filtering with the epidermal pigment absorption peak at about 680 nm as a reference to subtract the background. The Raman data is subjected to polynomial baseline correction to eliminate the fluorescence background. Feature extraction and matching: Extract the intensity of characteristic peaks from the processed spectra of each band and match them with a preset standard spectral library; Fusion Quantification: The intensity of the characteristic peaks obtained from feature extraction and matching is input into the PLS or ANN fusion model for quantitative analysis, outputting the type and concentration of each target pesticide in the sample and generating a detection report; Cleaning and data backup: After the test is completed, rinse the sample cell and probe surface with deionized water and wipe them dry. Back up the original spectral files and test reports for historical traceability and retesting.

7. The in-situ determination method for pesticide residue concentration in fruits and vegetables according to claim 6, characterized in that: In the system calibration and SNR adjustment, a pesticide standard sample with a known concentration is used as the calibration sample. The spectral signal of the standard sample is acquired by the CMOS imaging module. The computer software calculates the signal-to-noise ratio based on the spectral signal. By adjusting the light source intensity and the adjustable slit width, the signal-to-noise ratio is made not less than 30dB. After the signal-to-noise ratio is not less than 30dB, the module selection and timing acquisition are started.

8. The in-situ determination method for pesticide residue concentration in fruits and vegetables according to claim 7, characterized in that: In the module selection and timing acquisition, for organophosphorus pesticides, a combination of ultraviolet-visible spectrophotometry, near-infrared spectroscopy, and enhanced Raman probe is used for acquisition; for pyrethroid pesticides, a combination of fluorescence excitation, near-infrared spectroscopy, and enhanced Raman probe is used for acquisition. The ultraviolet-visible spectrophotometry and fluorescence excitation modules are acquired alternately at a time interval of 0.5 seconds, while the near-infrared spectroscopy and enhanced Raman probe are acquired synchronously within the same time window via fiber optic splitter.

9. The in-situ determination method for pesticide residue concentration in fruits and vegetables according to claim 8, characterized in that: In the preprocessing and background subtraction, baseline correction and intensity normalization are first performed on the UV-Vis and near-infrared spectral data; then, a background reference is constructed based on the epidermal pigment absorption peak at 680 nm in the UV-Vis reflectance spectrum, and adaptive filtering is performed on the fluorescence spectral data; subsequently, polynomial baseline correction is performed on the Raman spectral data; after the above processing is completed, the intensity of characteristic peaks is extracted from the processed UV-Vis, near-infrared, fluorescence, and Raman spectra.

10. The in-situ determination method for pesticide residue concentration in fruits and vegetables according to claim 9, characterized in that: In the fusion quantification, the intensity of each band feature peak obtained through the feature extraction and matching steps is used as input and input to a partial least squares regression model or an artificial neural network model for quantitative calculation. The artificial neural network model includes an input layer, three hidden layers and an output layer, and each output node of the output layer corresponds to the concentration of each target pesticide in the sample.

Images