Rapid detection method for heavy metals in agricultural products based on spectral technology

Abstract

This invention relates to the fields of analytical detection and quantum dot fluorescence technology, and discloses a rapid detection method for heavy metals in agricultural products based on spectroscopic technology. The method includes: acquiring a set of characteristic parameters of the clarified extract of agricultural products and its matrix; generating correction coefficients by calling a matrix interference compensation function; acquiring three-dimensional fluorescence spectra after reacting with a composite quantum dot probe; analyzing the spectral data using a multi-channel convolutional neural network model to output a preliminary concentration estimate; and obtaining the final heavy metal concentration by combining the correction coefficients and determining whether it exceeds the standard. This invention achieves simultaneous detection of multiple elements such as lead, cadmium, and mercury, possessing high sensitivity, strong generalization, and field applicability, with detection limits down to the microgram / kilogram level, and low processing time.

Description

Technical Field

[0001] This invention belongs to the field of analytical detection and quantum dot fluorescence technology, specifically relating to a rapid detection method for heavy metals in agricultural products based on spectral technology. Background Technology

[0002] With increasingly stringent food safety regulations, rapid and accurate detection of heavy metal contamination in agricultural products has become a crucial step in safeguarding public health. Spectroscopic techniques, due to their non-destructive nature, high sensitivity, and real-time response, have been widely applied in the detection of heavy metals in agricultural products. Quantum dot fluorescent probe-based detection methods identify target contaminants by leveraging the specific interaction between heavy metal ions and ligands on the quantum dot surface to induce changes in fluorescence signals.

[0003] However, this technology has the following problems in practical applications: the design of quantum dot probes is highly dependent on empirical trial and error and lacks systematic guiding principles, resulting in long screening cycles and high costs; at the same time, different agricultural product matrices vary significantly in terms of compositional complexity, pH value, and organic matter content, resulting in poor response stability of the same probe to different samples and serious cross-interference, which severely restricts the generalization ability and field applicability of the detection method.

[0004] The core of spectral response-based heavy metal detection methods lies in constructing highly selective and sensitive quantum dot-matrix adaptation systems. An ideal probe needs to maintain stable optical properties in complex agricultural product extracts and produce distinguishable fluorescence quenching or enhancement effects for specific heavy metal ions. However, existing technologies mostly employ a single type of quantum dot combined with fixed ligands for detection, making it difficult to cover diverse matrix environments and multi-metal coexistence scenarios.

[0005] In existing technologies, the optimization of quantum dot probes mainly relies on the artificial synthesis of a large number of candidate materials and the testing of their performance in standard solutions one by one before attempting to transfer them to real samples. This process is inefficient and cannot predict the influence of matrix effects. Even with the introduction of some machine learning models to assist in prediction, the lack of interpretability of these models makes it difficult to guide the rational design of probe molecules. Although microfluidic chips can achieve sample pretreatment and detection integration, their channel structure and quantum dot array layout are not optimized in synergy with the probe response mechanism, resulting in limited detection throughput and reliability.

[0006] The aforementioned problems collectively make it difficult for existing spectral detection methods to achieve rapid and customized heavy metal screening on a single platform that can adapt to multiple matrices when dealing with various types of agricultural products. There is an urgent need for a new probe design paradigm that integrates high-throughput experiments, interpretable artificial intelligence, and intelligent optimization algorithms. Summary of the Invention

[0007] This invention provides a rapid detection method for heavy metals in agricultural products based on spectral technology, aiming to solve the technical problems of existing quantum dot probes, such as reliance on trial and error in design, low screening efficiency, and weak generalization ability to adapt to different agricultural product matrices. The method achieves high sensitivity, high selectivity, and cross-matrix generalization detection of multiple heavy metal ions by constructing a physicochemical response model of heavy metal ion-quantum dot interaction and combining it with a decoupling mechanism between agricultural product matrix characteristic parameters and spectral signals.

[0008] This invention provides a rapid detection method for heavy metals in agricultural products based on spectroscopic technology, comprising: Collect samples of agricultural products to be tested and pre-treat them to obtain a clear liquid extract; Obtain the matrix characteristic parameter set of the clarified liquid phase extract; Based on the matrix feature parameter set, a pre-constructed matrix interference compensation function is invoked to generate a spectral signal correction coefficient vector for the current agricultural product sample. The clarified liquid extract is mixed with a pre-set composite quantum dot probe solution, which contains at least three quantum dot components with different surface ligand structures and band gaps. Each quantum dot component has a specific fluorescence quenching response to lead ions, cadmium ions, and mercury ions, respectively. After reacting for a predetermined time under constant temperature and light-protected conditions, the mixed solution is irradiated with an excitation light source, and its three-dimensional fluorescence spectral data cube is collected simultaneously. The three-dimensional fluorescence spectral data cube includes the excitation wavelength dimension, the emission wavelength dimension, and the fluorescence intensity dimension. The three-dimensional fluorescence spectral data cube is input into a pre-trained multi-channel convolutional neural network model. The multi-channel convolutional neural network model contains three parallel feature extraction branches, each corresponding to a target heavy metal ion. Each branch uses a depthwise separable convolutional structure to extract local spectral fingerprint features and integrates global contextual information through an attention weight fusion mechanism. The multi-channel convolutional neural network model outputs a preliminary vector of heavy metal concentration estimates. The initial heavy metal concentration estimate vector is multiplied element-wise with the spectral signal correction coefficient vector to obtain the corrected final heavy metal concentration estimate vector. Based on the corrected final estimate vector of heavy metal concentration, it is determined whether each heavy metal ion in the agricultural product sample exceeds the national food safety limit standard.

[0009] Preferably, the matrix characteristic parameter set includes organic acid concentration, polyphenol content, protein mass concentration, molar concentration of carbohydrates, and turbidity value.

[0010] Preferably, obtaining the matrix characteristic parameter set of the clarified liquid phase extract includes: The concentrations of organic acids and the molar concentrations of carbohydrates were determined by high performance liquid chromatography. The polyphenol content was determined using the Folin-Ciocalteu method. Protein concentration was determined using the Coomassie Brilliant Blue method. The turbidity value was determined at a wavelength of 600 nm using a UV-Vis spectrophotometer.

[0011] Preferably, based on the matrix feature parameter set, a pre-constructed matrix interference compensation function is invoked to generate a spectral signal correction coefficient vector for the current agricultural product sample, including: Collect no fewer than 100 representative agricultural product samples, covering four major categories: leafy vegetables, root vegetables, fruits, and grains; Standardized pretreatment was performed on each type of sample to obtain the corresponding clear liquid phase extract; The concentrations of organic acids, polyphenols, protein mass concentration, molar concentration of carbohydrates, and turbidity of each clear liquid extract were determined to form a matrix characteristic parameter set. A spiked sample set was formed by adding known concentration gradients of lead ion, cadmium ion, and mercury ion standard solutions to each clear liquid phase extract; Fluorescence response tests were performed on all spiked samples using a standardized composite quantum dot probe solution, and the original fluorescence quenching rate was recorded. A multivariate nonlinear regression model was established with the matrix feature parameter set as input and the fluorescence quenching rate offset as output. The parameters were optimized using the support vector regression algorithm, and finally the matrix interference compensation function was formed.

[0012] Preferably, the preparation process of the composite quantum dot probe solution includes: Zinc cadmium sulfide core-shell quantum dots were selected as the basic luminescent material; The basic luminescent material was surface functionalized by grafting three ligand molecules: mercaptoacetic acid, sodium dimercaptopropanesulfonate, and glutathione, to form three quantum dot derivatives with different surface charge densities and steric hindrance effects. The three quantum dot derivatives were mixed in a phosphate buffer solution at a mass ratio of 1:1:1 to obtain the composite quantum dot probe solution.

[0013] Preferably, after mixing the clarified liquid extract with a pre-set composite quantum dot probe solution, the mixture is reacted under constant temperature and light-protected conditions for a predetermined time, including: The mixed solution was placed in a constant temperature and dark environment to allow the heavy metal ions to fully bind with the quantum dot surface ligands and reach fluorescence quenching equilibrium.

[0014] Preferably, the cube containing its three-dimensional fluorescence spectral data includes: Use a xenon lamp as the excitation source; The emission wavelength scanning range is 400 nm to 650 nm, with a step size of 2 nm; The fluorescence intensity sampling time for each excitation-emission combination is 0.5 seconds; The entire scanning process was completed in a constant temperature environment of 25 degrees Celsius.

[0015] Preferably, inputting the three-dimensional fluorescence spectral data cube into a pre-trained multi-channel convolutional neural network model includes: Construct a training dataset, in which each sample group consists of a three-dimensional fluorescence spectral data cube and its corresponding heavy metal real concentration label; Initialize the convolutional kernel weights for 3 parallel feature extraction branches, each branch containing 5 convolutional layers; A batch normalization layer and a modified linear unit activation function are added after each convolutional layer; A global average pooling layer is added after the 5th convolutional layer to compress the spatial dimension into a one-dimensional feature vector; The one-dimensional feature vectors output from the three branches are concatenated and then input into the fully connected layer, which contains two layers of hidden nodes. The output layer uses a linear activation function to directly output the predicted concentration values ​​of the three heavy metal ions; The mean squared error is used as the loss function, and the adaptive moment estimator optimizer is used to iteratively update the model parameters until the loss function value on the validation set converges.

[0016] Preferably, the preliminary heavy metal concentration estimate vector is multiplied element-wise with the spectral signal correction coefficient vector to obtain the corrected final heavy metal concentration estimate vector, including: Calculate correction factors for lead ions, cadmium ions, and mercury ions respectively. , For the first The offset of fluorescence quenching rate corresponding to each heavy metal; Multiply the initial concentration estimate by the corresponding correction factor to obtain the corrected final concentration estimate.

[0017] Preferably, determining whether each heavy metal ion in the tested agricultural product sample exceeds the national food safety limit standard based on the corrected final estimated value vector of heavy metal concentration includes: Compare the corrected lead ion concentration with the limit of 0.1 mg per kilogram; Compare the corrected cadmium ion concentration with the limit of 0.05 mg per kilogram; Compare the corrected mercury ion concentration with the limit of 0.01 mg per kilogram; If the concentration of any heavy metal exceeds its corresponding limit, the agricultural product sample is deemed unqualified.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a composite quantum dot probe system, which overcomes the limitations of traditional single quantum dot probes for specific heavy metal ions and achieves simultaneous response to three high-risk heavy metals: lead, cadmium, and mercury. 2. By introducing a matrix feature parameter set and a matrix interference compensation function, the non-specific interference of complex matrices in agricultural products on fluorescence signals is effectively decoupled, improving the generalization ability of the detection method across different types of agricultural products. A multi-channel convolutional neural network model is used to learn the three-dimensional fluorescence spectral data end-to-end, avoiding the subjectivity and inefficiency of traditional methods that rely on empirical formulas or manual feature extraction. 3. The entire testing process does not require complex sample digestion or large-scale instruments and equipment, the testing time can be controlled, and the detection limit reaches the microgram per kilogram level, meeting the actual needs of rapid on-site screening. 4. This method does not rely on trial-and-error probe screening. The probe formulation is fixed and can be prepared in batches, which greatly reduces the detection cost and operation threshold, and provides reliable technical support for the supervision of agricultural product quality and safety. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the composite quantum dot probe and the specific response of heavy metal ions and the matrix interference compensation mechanism in this invention; Figure 3 This is a flowchart illustrating the logical process framework for the preprocessing of agricultural product samples and the construction of matrix feature parameter sets in this invention. Figure 4 This is a flowchart illustrating the logical flow of the three-dimensional fluorescence spectral data cube acquisition and multi-channel convolutional neural network concentration analysis in this invention. Figure 5 This is a flowchart illustrating the logical process of constructing the matrix interference compensation function and generating the correction coefficients in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow of each functional unit of the terminal detection system in this invention. Detailed Implementation

[0020] refer to Figures 1 to 6This invention provides a rapid detection method for heavy metals in agricultural products based on spectral technology, aiming to solve the technical problems of existing quantum dot probes, such as reliance on trial and error in design, low screening efficiency, and weak generalization ability to adapt to different agricultural product matrices. This method constructs a physicochemical response model of heavy metal ion-quantum dot interaction, combined with a decoupling mechanism between agricultural product matrix characteristic parameters and spectral signals, to achieve simultaneous detection of three high-risk heavy metals: lead, cadmium, and mercury ions, and possesses high generalization ability across agricultural product types.

[0021] The method includes the following steps: Collect samples of agricultural products to be tested and pre-treat them to obtain a clear liquid extract; Obtain a matrix characteristic parameter set of the clarified liquid extract, the matrix characteristic parameter set including organic acid concentration, polyphenol content, protein mass concentration, molar concentration of carbohydrates, and turbidity value; Based on the matrix feature parameter set, a pre-constructed matrix interference compensation function is invoked to generate a spectral signal correction coefficient vector for the current agricultural product sample. The clarified liquid extract is mixed with a pre-set composite quantum dot probe solution, which contains at least three quantum dot components with different surface ligand structures and band gaps. Each quantum dot component has a specific fluorescence quenching response to lead ions, cadmium ions, and mercury ions, respectively. After reacting for a predetermined time under constant temperature and light-protected conditions, the mixed solution is irradiated with an excitation light source, and its three-dimensional fluorescence spectral data cube is collected simultaneously. The three-dimensional fluorescence spectral data cube includes the excitation wavelength dimension, the emission wavelength dimension, and the fluorescence intensity dimension. The three-dimensional fluorescence spectral data cube is input into a pre-trained multi-channel convolutional neural network model. The multi-channel convolutional neural network model contains three parallel feature extraction branches, each corresponding to a target heavy metal ion. Each branch uses a depthwise separable convolutional structure to extract local spectral fingerprint features and integrates global contextual information through an attention weight fusion mechanism. The multi-channel convolutional neural network model outputs a preliminary vector of heavy metal concentration estimates. The initial heavy metal concentration estimate vector is multiplied element-wise with the spectral signal correction coefficient vector to obtain the corrected final heavy metal concentration estimate vector. Based on the corrected final estimate vector of heavy metal concentration, it is determined whether each heavy metal ion in the agricultural product sample exceeds the national food safety limit standard.

[0022] First, perform step S1: collect samples of the agricultural products to be tested and pre-treat them to obtain a clear liquid extract.

[0023] Specifically, 10 grams of the agricultural product sample to be tested was placed in a stainless steel grinding jar, and sufficient liquid nitrogen was injected to completely freeze and embrittle it. Then, a high-frequency vibrating ball mill was used for freeze grinding for 3 minutes to ensure that the sample was crushed into uniform powder with a particle size of less than 100 micrometers.

[0024] The obtained powder was transferred to a 50 mL polypropylene centrifuge tube, and 20 mL of a methanol aqueous solution containing 0.1% trifluoroacetic acid was added. The solvent system consisted of methanol and ultrapure water in a volume ratio of 7:3. Trifluoroacetic acid was weighed precisely to a final concentration of 1 / 1000 using an analytical balance.

[0025] Place the centrifuge tubes on a vortex mixer and vortex at 2500 rpm for 5 minutes to allow the target heavy metal ions to dissolve fully from the agricultural product matrix. After vortexing, place the centrifuge tubes in a high-speed refrigerated centrifuge and centrifuge at 4 degrees Celsius and 10000 rpm for 10 minutes to allow the insoluble solid particles to settle to the bottom of the tube.

[0026] Carefully aspirate the supernatant, avoiding disturbing the precipitate, and transfer it to another clean centrifuge tube. Then, vacuum filter the supernatant through a 0.22-micron PTFE membrane pre-wetted with ultrapure water to minimize adsorption loss. The filtrate, the clear liquid extract, should be sealed and stored at 4°C in the dark for use in subsequent steps.

[0027] Next, step S2 is performed: The matrix characteristic parameter set of the clarified liquid extract is obtained. This matrix characteristic parameter set includes organic acid concentration, polyphenol content, protein mass concentration, carbohydrate molar concentration, and turbidity value. The organic acid concentration is determined using high-performance liquid chromatography (HPLC) with a C18 reversed-phase column, a gradient elution system of phosphate buffer and acetonitrile, a flow rate of 1 mL / min, a detection wavelength of 210 nm, and external standard method for quantification.

[0028] The molar concentration of carbohydrates was also determined by high-performance liquid chromatography (HPLC), but an amino column was used as the stationary phase, and an evaporative light scattering detector was employed. Calibration curves were established using glucose, fructose, and sucrose standards. The polyphenol content was determined using the Folin-Ciocalteu method: 0.5 mL of the clear HPLC extract was added to 2.5 mL of Folin-Ciocalteu reagent, mixed, and allowed to stand for 5 minutes. Then, 2 mL of 7.5% sodium carbonate solution was added, and the mixture was reacted in a 37°C water bath for 60 minutes. The absorbance was measured at 760 nm, and a standard curve was plotted using gallic acid as a standard.

[0029] Protein concentration was determined using the Coomassie Brilliant Blue method: 0.1 mL of sample was mixed with 5 mL of Coomassie Brilliant Blue G250 staining solution, allowed to stand at room temperature for 5 minutes, and the absorbance was measured at a wavelength of 595 nm. Bovine serum albumin was used as a standard for quantification.

[0030] Turbidity values ​​were determined directly at a wavelength of 600 nm using a UV-Vis spectrophotometer, measuring the absorbance of the clear liquid extract. This absorbance value served as the turbidity indicator, requiring no additional calibration. All measurements were performed within two hours of sample preparation to prevent component degradation from affecting parameter accuracy.

[0031] Then, step S3 is executed: based on the matrix feature parameter set, the pre-constructed matrix interference compensation function is called to generate a spectral signal correction coefficient vector for the current agricultural product sample.

[0032] The construction process of the matrix interference compensation function is as follows: collect no less than 100 representative agricultural product samples, covering four major categories: leafy vegetables such as spinach and lettuce, root vegetables such as carrots and potatoes, fruits such as apples and tomatoes, and grains such as rice and wheat.

[0033] Standardized pretreatment was performed on each type of sample to obtain the corresponding clear liquid extract. The concentrations of organic acids, polyphenols, protein mass concentration, molar concentration of carbohydrates, and turbidity were measured for each clear liquid extract to form a matrix characteristic parameter set. Lead, cadmium, and mercury ion standard solutions with known concentration gradients of 0, 5, 10, 20, and 50 μg / kg were added to each clear liquid extract to form a spiked sample set.

[0034] Fluorescence response tests were performed on all spiked samples using a standardized composite quantum dot probe solution, and the original fluorescence quenching rate was recorded. The fluorescence quenching rate was defined as the difference between the fluorescence intensity of the unspiked sample and the fluorescence intensity of the spiked sample, divided by the fluorescence intensity of the unspiked sample. A multivariate nonlinear regression model was established with the matrix characteristic parameter set as input and the fluorescence quenching rate offset as output.

[0035] The fluorescence quenching rate offset refers to the difference between the actual measured quenching rate and the theoretical interference-free quenching rate. Support vector regression was used for parameter optimization, with a radial basis function chosen as the kernel function. The penalty coefficient C and kernel parameter γ were determined to be optimal through five-fold cross-validation. The final matrix interference compensation function expression is as follows: ; This represents the fluorescence quenching rate offset. This refers to the concentration of organic acids. Polyphenol content, This refers to the protein mass concentration. This refers to the molar concentration of carbohydrates. This represents the turbidity value. For the current sample to be tested, its matrix characteristic parameters are substituted into this function to calculate the offsets corresponding to the three heavy metals. The first spectral signal correction coefficient vector Each element is defined , for the first Each fluorescence quenching rate offset is used to generate a three-dimensional correction coefficient vector. .

[0036] Perform step S4: Mix the clarified liquid extract with a pre-prepared composite quantum dot probe solution. The preparation process of the composite quantum dot probe solution includes: selecting zinc cadmium sulfide core-shell structured quantum dots as the basic luminescent material, with a particle size range of 2.5 nm to 4.5 nm, a half-width at half-maximum (WHM) of less than 30 nm, and a quantum yield greater than 70%.

[0037] The basic luminescent material was surface functionalized as follows: excess mercaptoacetic acid was added to the first quantum dot dispersion, and the mixture was stirred at 60 degrees Celsius for 12 hours under nitrogen protection to form negatively charged carboxyl-terminated quantum dots; sodium dimercaptopropanesulfonate was added to the second quantum dot dispersion, and the mixture was reacted under the same conditions to form sulfonate-modified quantum dots with greater steric hindrance; glutathione was added to the third quantum dot dispersion, and the mixture was stirred at room temperature in the dark for 8 hours to form bifunctional ligand-modified quantum dots containing thiol and amino groups.

[0038] Three modified quantum dot derivatives were purified by dialysis, and their concentrations were determined. They were then mixed in a phosphate buffer solution at a mass ratio of 1:1:1 (10 mmol / L), with the pH adjusted to 7.4, to obtain a composite quantum dot probe solution with a total quantum dot concentration of 0.5 mmol / L. 0.5 mL of the clear liquid extract and 0.5 mL of the composite quantum dot probe solution were mixed in a brown glass cuvette and gently shaken.

[0039] Perform step S5: After reacting for a predetermined time under constant temperature and light-protected conditions, irradiate the mixed solution with an excitation light source and simultaneously acquire its three-dimensional fluorescence spectral data cube. The reaction conditions are set at 25 degrees Celsius, in the dark, and allowed to stand for 15 minutes to ensure that the heavy metal ions fully bind to the quantum dot surface ligands and reach fluorescence quenching equilibrium.

[0040] The acquisition process of the three-dimensional fluorescence spectroscopy data cube includes: using a 300-watt xenon lamp as the excitation source, with an excitation wavelength scanning range of 250 nm to 450 nm and a step size of 5 nm, for a total of 41 excitation wavelength points; and an emission wavelength scanning range of 400 nm to 650 nm and a step size of 2 nm, for a total of 126 emission wavelength points; the fluorescence intensity sampling time for each excitation-emission combination is 0.5 seconds, and the photomultiplier tube voltage is set to 700 volts; the entire scanning process is completed in a dark room at a constant temperature of 25 degrees Celsius, and the slit width of the monochromator is fixed at 5 nm to ensure a balance between signal-to-noise ratio and resolution.

[0041] After data acquisition, a three-dimensional data cube with dimensions of 41×126×1 is formed. Each layer corresponds to the total emission spectrum at an excitation wavelength, and the values ​​are relative fluorescence intensities.

[0042] Step S6: Input the three-dimensional fluorescence spectroscopy data cube into a pre-trained multi-channel convolutional neural network model. The training process of the multi-channel convolutional neural network model includes: constructing a training dataset containing 5200 samples, each sample consisting of a three-dimensional fluorescence spectroscopy data cube and its corresponding heavy metal true concentration label, the label being determined by inductively coupled plasma mass spectrometry to ensure accuracy; initializing the convolutional kernel weights of three parallel feature extraction branches, each branch having the same structure and containing 5 convolutional layers.

[0043] The first convolutional layer uses 32 3×3 convolutional kernels, the second layer uses 64 5×5 convolutional kernels, the third layer uses 64 3×3 convolutional kernels, the fourth layer uses 128 7×7 convolutional kernels, and the fifth layer uses 128 3×3 convolutional kernels. All convolutional operations use same padding with a stride of 1.

[0044] A batch normalization layer and a modified linear unit activation function are applied after each convolutional layer. A global average pooling layer is applied after the 5th convolutional layer to compress the spatial dimension into a one-dimensional feature vector of length 128. Three branches process the same input data cube but learn the spectral response patterns of different heavy metals. The one-dimensional feature vectors output from the three branches are concatenated into a joint feature vector of length 384, which is then input into the fully connected layer.

[0045] The fully connected layer contains two layers of hidden nodes: the first layer has 256 nodes, and the second layer has 64 nodes, both activated using modified linear units (MLUs). The output layer has three neurons, using a linear activation function, and directly outputs the predicted concentrations of lead, cadmium, and mercury (in micrograms per kilogram). Mean squared error is used as the loss function. : ; The total number of samples participating in the loss calculation. For the first The true value of lead (Pb) concentration in each sample. For the first The model-predicted values ​​of lead (Pb) concentration in each sample. For the first The true value of cadmium (Cd) concentration in each sample. For the first Model-predicted values ​​of cadmium (Cd) concentration in a sample. For the first The true value of mercury (Hg) concentration in each sample. For the first Model-predicted values ​​of mercury (Hg) concentration in a sample.

[0046] An adaptive moment estimator (AME) was used to iteratively update the model parameters. The initial learning rate was 0.001, the batch size was 32, and the training run consisted of 200 epochs. The training was terminated early when the validation set loss did not decrease for 10 consecutive epochs. After training, the model was deployed on an embedded industrial control computer platform, receiving real-time acquired 3D fluorescence spectral data cubes and outputting a vector of preliminary heavy metal concentration estimates. . This is a preliminary estimate of the lead (Pb) concentration directly output by the model. This is a preliminary estimate of the cadmium (Cd) concentration directly output by the model. This is a preliminary estimate of the mercury (Hg) concentration directly output by the model.

[0047] Perform step S7: Multiply the preliminary heavy metal concentration estimate vector element-wise with the spectral signal correction coefficient vector to obtain the corrected final heavy metal concentration estimate vector. That is: ; ; ; This is the final estimate of lead (Pb) concentration after matrix correction. This represents the spectral signal correction coefficient for lead (Pb). This is the final estimated cadmium (Cd) concentration after matrix correction. This represents the spectral signal correction coefficient for cadmium (Cd). This is the final estimated mercury (Hg) concentration after matrix correction. This is the spectral signal correction coefficient for mercury (Hg).

[0048] The correction process effectively compensates for the non-specific enhancement or suppression of fluorescence signals caused by differences in agricultural product matrix, making the concentration estimate closer to the true value.

[0049] Finally, step S8 is executed: based on the corrected final estimated value vector of heavy metal concentration, it is determined whether each heavy metal ion in the tested agricultural product sample exceeds the national food safety limits. The national food safety limits are implemented in accordance with the "National Food Safety Standard: Limits of Contaminants in Food". The limits for lead are 0.01 mg / kg, for cadmium 0.05 mg / kg, and for mercury 0.01 mg / kg.

[0050] The corrected concentration values ​​are compared with their corresponding limits. If any heavy metal concentration exceeds its limit, the agricultural product sample is deemed unqualified, and the message "Exceeds Limit" along with the specific element and its concentration value is displayed on the human-machine interface. If all values ​​are below the limit, "Qualified" is displayed. The results are simultaneously stored in a local database and can be uploaded to the regulatory platform via a wireless module.

[0051] The system upon which the above method relies includes a sample pretreatment unit, a matrix parameter analysis unit, a composite quantum dot probe supply unit, a spectral signal acquisition unit, a data processing and concentration analysis unit, and a result determination unit. The sample pretreatment unit consists of a liquid nitrogen cryogenic grinder, a vortex oscillator, a high-speed centrifuge, and a vacuum filtration device, realizing the automated preparation from raw agricultural products to a clear liquid-phase extract.

[0052] The matrix parameter analysis unit integrates a high-performance liquid chromatograph, a UV-Vis spectrophotometer, and matching reagent kits to rapidly determine five matrix parameters.

[0053] The composite quantum dot probe supply unit includes a temperature-controlled, light-proof storage tank, a precision peristaltic pump, and a mixing chamber to ensure stable storage and precise release of the probe solution.

[0054] The spectral signal acquisition unit uses a 300-watt xenon lamp as the excitation source and is equipped with a dual monochromator system and a high-sensitivity photomultiplier tube. The wavelength accuracy is better than 0.5 nanometers and the dark current is less than 1.5 nanoamps, ensuring high-quality acquisition of three-dimensional fluorescence spectral data.

[0055] The data processing and concentration analysis unit runs on an embedded industrial control computer platform, equipped with 16 gigabytes of memory and a solid-state drive. The operating system is a real-time Linux kernel, and it has a built-in matrix interference compensation function module and a multi-channel convolutional neural network model module to perform all computational tasks.

[0056] The result determination unit outputs the determination results through a graphical user interface and supports printing and data export. All units are interconnected via an industrial bus, and the data flow follows a strict time synchronization protocol, ensuring the entire testing process is completed within 30 minutes, with a detection limit of 0.5 micrograms per kilogram, meeting the needs of rapid on-site screening.

[0057] This embodiment fully realizes the entire process from sample collection to result determination, solving the core defects of traditional methods such as probe screening relying on trial and error, severe matrix interference, and poor cross-category generalization ability. By combining a composite quantum dot probe with a fixed formulation, a matrix interference compensation function constructed based on measured data, and an end-to-end trained multi-channel convolutional neural network model, it achieves rapid detection of heavy metals with high sensitivity, high selectivity, and strong generalization.

[0058] 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.

[0059] 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. A rapid detection method for heavy metals in agricultural products based on spectroscopic technology, characterized in that, include: Collect samples of agricultural products to be tested and pre-treat them to obtain a clear liquid extract; Obtain the matrix characteristic parameter set of the clarified liquid phase extract; Based on the matrix feature parameter set, a pre-constructed matrix interference compensation function is invoked to generate a spectral signal correction coefficient vector for the current agricultural product sample. The clarified liquid extract is mixed with a pre-set composite quantum dot probe solution, which contains at least three quantum dot components with different surface ligand structures and band gaps. Each quantum dot component has a specific fluorescence quenching response to lead ions, cadmium ions, and mercury ions, respectively. After reacting for a predetermined time under constant temperature and light-protected conditions, the mixed solution is irradiated with an excitation light source, and a three-dimensional fluorescence spectral data cube is collected simultaneously. The three-dimensional fluorescence spectral data cube includes an excitation wavelength dimension, an emission wavelength dimension, and a fluorescence intensity dimension. The three-dimensional fluorescence spectral data cube is input into a pre-trained multi-channel convolutional neural network model. The multi-channel convolutional neural network model contains three parallel feature extraction branches, each corresponding to a target heavy metal ion. Each branch uses a depthwise separable convolutional structure to extract local spectral fingerprint features and integrates global contextual information through an attention weight fusion mechanism. The multi-channel convolutional neural network model outputs a preliminary vector of heavy metal concentration estimates. The initial heavy metal concentration estimate vector is multiplied element-wise with the spectral signal correction coefficient vector to obtain the corrected final heavy metal concentration estimate vector. Based on the corrected final estimate vector of heavy metal concentration, it is determined whether each heavy metal ion in the agricultural product sample exceeds the national food safety limit standard.

2. The rapid detection method for heavy metals in agricultural products based on spectral technology according to claim 1, characterized in that, The matrix characteristic parameter set includes organic acid concentration, polyphenol content, protein mass concentration, molar concentration of carbohydrates, and turbidity value.

3. The rapid detection method for heavy metals in agricultural products based on spectral technology according to claim 2, characterized in that, Obtaining the matrix characteristic parameter set of the clarified liquid phase extract includes: The concentrations of organic acids and the molar concentrations of carbohydrates were determined by high performance liquid chromatography. The polyphenol content was determined using the Folin-Ciocalteu method. Protein concentration was determined using the Coomassie Brilliant Blue method. The turbidity value was determined at a wavelength of 600 nm using a UV-Vis spectrophotometer.

4. The rapid detection method for heavy metals in agricultural products based on spectral technology according to claim 3, characterized in that, Based on the matrix feature parameter set, a pre-constructed matrix interference compensation function is invoked to generate a spectral signal correction coefficient vector for the current agricultural product sample, including: Collect no fewer than 100 representative agricultural product samples, covering four major categories: leafy vegetables, root vegetables, fruits, and grains; Standardized pretreatment was performed on each type of sample to obtain the corresponding clear liquid phase extract; The concentrations of organic acids, polyphenols, protein mass concentration, molar concentration of carbohydrates, and turbidity of each clear liquid extract were determined to form a matrix characteristic parameter set. A spiked sample set was formed by adding known concentration gradients of lead ion, cadmium ion, and mercury ion standard solutions to each clear liquid phase extract; Fluorescence response tests were performed on all spiked samples using a standardized composite quantum dot probe solution, and the original fluorescence quenching rate was recorded. A multivariate nonlinear regression model was established with the matrix feature parameter set as input and the fluorescence quenching rate offset as output. The parameters were optimized using the support vector regression algorithm, and finally the matrix interference compensation function was formed.

5. The rapid detection method for heavy metals in agricultural products based on spectral technology according to claim 4, characterized in that, The preparation process of the composite quantum dot probe solution includes: Zinc cadmium sulfide core-shell quantum dots were selected as the basic luminescent material; The basic luminescent material was surface functionalized by grafting three ligand molecules: mercaptoacetic acid, sodium dimercaptopropanesulfonate, and glutathione, to form three quantum dot derivatives with different surface charge densities and steric hindrance effects. The three quantum dot derivatives were mixed in a phosphate buffer solution at a mass ratio of 1:1:1 to obtain the composite quantum dot probe solution.

6. The rapid detection method for heavy metals in agricultural products based on spectral technology according to claim 5, characterized in that, The clarified liquid extract is mixed with a pre-set composite quantum dot probe solution and reacted under constant temperature and light-protected conditions for a predetermined time, including: The mixed solution was placed in a constant temperature and dark environment to allow the heavy metal ions to fully bind with the quantum dot surface ligands and reach fluorescence quenching equilibrium.

7. The rapid detection method for heavy metals in agricultural products based on spectral technology according to claim 6, characterized in that, The cube for acquiring three-dimensional fluorescence spectral data includes: Use a xenon lamp as the excitation source; The emission wavelength scanning range is 400 nm to 650 nm, with a step size of 2 nm; The fluorescence intensity sampling time for each excitation-emission combination is 0.5 seconds; The entire scanning process was completed in a constant temperature environment of 25 degrees Celsius.

8. The rapid detection method for heavy metals in agricultural products based on spectral technology according to claim 7, characterized in that, The three-dimensional fluorescence spectral data cube is input into a pre-trained multi-channel convolutional neural network model, including: Construct a training dataset, in which each sample group consists of a three-dimensional fluorescence spectral data cube and its corresponding heavy metal real concentration label; Initialize the convolutional kernel weights for 3 parallel feature extraction branches, each branch containing 5 convolutional layers; A batch normalization layer and a modified linear unit activation function are added after each convolutional layer; A global average pooling layer is added after the 5th convolutional layer to compress the spatial dimension into a one-dimensional feature vector; The one-dimensional feature vectors output from the three branches are concatenated and then input into the fully connected layer, which contains two layers of hidden nodes. The output layer uses a linear activation function to directly output the predicted concentration values ​​of the three heavy metal ions; The mean squared error is used as the loss function, and the adaptive moment estimator optimizer is used to iteratively update the model parameters until the loss function value on the validation set converges.

9. The rapid detection method for heavy metals in agricultural products based on spectral technology according to claim 8, characterized in that, The initial heavy metal concentration estimate vector is multiplied element-wise with the spectral signal correction coefficient vector to obtain the corrected final heavy metal concentration estimate vector, including: Calculate correction factors for lead ions, cadmium ions, and mercury ions respectively. , For the first The offset of fluorescence quenching rate corresponding to each heavy metal; Multiply the initial concentration estimate by the corresponding correction factor to obtain the corrected final concentration estimate.

10. The rapid detection method for heavy metals in agricultural products based on spectroscopic technology according to claim 9, characterized in that, Based on the corrected final estimate vector of heavy metal concentrations, determine whether each heavy metal ion in the tested agricultural product sample exceeds the national food safety limit standard, including: Compare the corrected lead ion concentration with the limit of 0.1 mg per kilogram; Compare the corrected cadmium ion concentration with the limit of 0.05 mg per kilogram; Compare the corrected mercury ion concentration with the limit of 0.01 mg per kilogram; If the concentration of any heavy metal exceeds its corresponding limit, the agricultural product sample is deemed unqualified.

Images