Rapid ingredient detection system based on food additives

By constructing a multi-channel response detection system and combining feature fusion and matching analysis of spectral, fluorescence and electrochemical signals, the problem of rapid detection of food additives in complex matrix environments has been solved, achieving high-confidence food additive identification. It is suitable for mobile rapid testing and on-site law enforcement in various food scenarios.

CN121499756APending Publication Date: 2026-02-10滨州市检验检测中心(滨州市纺织纤维检验所) +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511683169.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-11-17
Publication Date
2026-02-10

AI Technical Summary

Technical Problem

Existing technologies struggle to detect food additives quickly and accurately in complex food matrix environments, especially in remote rural school canteens, small catering businesses, or night market stalls. Traditional detection methods are time-consuming, require expensive equipment, and are complex to operate, making rapid on-site screening impossible. Furthermore, they are prone to false positives or false negatives when multiple interfering factors are present.

Method used

A multi-channel response detection system driven by food matrix information is adopted, which combines feature fusion and matching analysis of spectral, fluorescence and electrochemical signals. By constructing additive fingerprint feature groups and matching degree scoring mechanism, rapid and intelligent identification of food additives can be achieved.

Benefits of technology

It achieves high-confidence rapid judgment in complex food matrix environments, breaks through the technical bottleneck of traditional detection methods, has broad spectrum and real-time performance, can identify unknown or unregistered illegal additives, and is suitable for mobile rapid testing and on-site law enforcement in various food scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121499756A_ABST
    Figure CN121499756A_ABST
Patent Text Reader

Abstract

The invention discloses a rapid component detection system based on food additives, and relates to the technical field of food safety detection.A food sample is collected, matrix information of the food sample is obtained, an additive fingerprint feature group is extracted based on a reference model, sample response signals are collected through multi-channel detection, and feature matching analysis is carried out; outputting additive existence judgment and a matching degree score; when the matching degree score is lower than a preset threshold value, abnormal response features are extracted, feature vectors of the components to be verified are constructed and compared with the high-risk component template data, and suspected illegal added components are marked; by fusing spectrum, fluorescence and electrochemical multi-dimensional detection channels, the method has the characteristics of broad spectrum, rapidness and high adaptability, can realize intelligent identification and early warning of unknown or illegal additives, and is suitable for mobile detection and risk screening in various food scenes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of food safety testing technology, and specifically to a rapid detection system based on food additive components. Background Technology

[0002] Food safety issues are receiving increasing attention. Food additives, as important components for improving food quality, extending shelf life, and enhancing taste, are widely used in various processed foods such as beverages, snacks, and condiments. However, some illegal manufacturers may add excessive or unlicensed additives, such as phthalate preservatives, sodium formaldehyde sulfoxylate, and formaldehyde sulfoxylate, to reduce costs, improve taste, or falsify expiration dates, posing serious food safety risks.

[0003] Current methods for detecting food additives mainly include liquid chromatography, mass spectrometry, and gas chromatography. While these methods offer high precision, they generally suffer from long testing cycles, expensive equipment, complex operation, and high technical requirements for testing personnel. This makes rapid on-site screening difficult, especially in challenging environments such as remote rural school canteens, small catering businesses, or night market vendors. The lack of rapid detection methods hinders the timely detection and handling of illegal additives.

[0004] Furthermore, food additives are diverse and their matrices vary greatly, making it difficult for traditional detection methods to achieve both broad-spectrum and specificity. Especially when multiple interfering factors (such as high fat, high salt, and high sugar) are present simultaneously, false positives or false negatives are prone to occur, reducing the accuracy of the test results. Summary of the Invention

[0005] The purpose of this invention is to provide a rapid detection system for food additive components to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a rapid detection system for food additive components, comprising: Sample collection module: Collects food samples to be tested and preprocesses them to obtain food matrix information F for analysis, where F includes the sample's pH value, viscosity, conductivity and color distribution parameters; Matrix modeling module: Based on the food matrix information F, it matches a preset food matrix model library to determine the reference model R that is closest to F; Additive feature extraction module: Based on the reference model R, extract the corresponding additive fingerprint feature group T from the additive feature response database, where T includes the predefined target additive's spectral absorption peak, fluorescence response value and electrochemical characteristic curve in the matrix; Multi-channel detection module: performs real-time multi-channel detection on the food sample and acquires the corresponding detection signal S′; Feature matching module: Performs feature matching analysis between the detection signal S′ and the fingerprint feature group T of the additive to obtain the judgment result of whether the target additive exists in the current sample, and outputs the matching score P; Anomaly detection module: When the matching score P is less than the preset confidence threshold, it is determined that there is an unknown additive in the sample, its abnormal features are extracted, and a feature vector of the component to be verified is constructed based on the abnormal features; Illegal ingredient identification module: compares the feature vector of the ingredient to be verified with the template data of high-risk ingredients, and marks suspected illegally added ingredients.

[0007] Preferably, the step of matching a preset food matrix model library with the food matrix information F to determine the reference model R closest to F includes: The parameters in the acquired food matrix information F are standardized to construct a matrix feature vector F′ under a unified dimension. The similarity between F′ and the feature vectors of each reference model in the preset food matrix model library is calculated using the Euclidean distance algorithm or the cosine similarity algorithm. Set a similarity threshold, and select the reference model R with the highest similarity that is greater than the threshold based on the calculation results.

[0008] Preferably, based on the reference model R, the corresponding additive fingerprint feature group T is extracted from the additive feature response database, including: The food matrix category information corresponding to the reference model R is input into the additive feature response database to limit the search scope and filter the target additive set that matches the matrix environment. Feature fusion is performed on the multi-source response data of each additive in the selected target additive set. The multi-source response data includes spectral absorption peaks, fluorescence response values, and electrochemical potential change curves. A dimensionality reduction algorithm based on principal component analysis was used to extract features from the fused multidimensional response data to obtain feature vectors characterizing the key response behaviors of each additive. The feature vectors are standardized and stored as additive fingerprint feature groups T, and a one-to-one correspondence is established with the reference model R.

[0009] Preferably, T includes the predefined spectral absorption peaks, fluorescence response values, and electrochemical characteristic curves of the target additive in the matrix, including: The matrix parameters corresponding to the reference model R are called, and the response regulation factor of the additive in this type of matrix environment is set to correct the baseline state of the additive response characteristics. Based on the regulatory factors, the original response data of each target additive in the database were adapted and transformed to obtain the response reconstruction results of the spectral absorption curve, fluorescence emission intensity curve and electrochemical potential change curve under this type of matrix. The main absorption peak position in the spectral curve is identified by the peak extraction algorithm, the maximum excitation value and corresponding emission value of the fluorescence response are extracted, and the main reduction / oxidation potential range in the electrochemical response is extracted to form a standard response feature subset.

[0010] Preferably, acquiring the corresponding detection signal S′ includes: The pretreated food samples were assigned to the spectral detection unit, fluorescence detection unit, and electrochemical detection unit, respectively, corresponding to various response dimensions in the additive fingerprint feature group T. In the spectral detection unit, a broadband light source is used to illuminate the sample and collect the reflected or transmitted signals to generate a spectral absorption curve within the wavelength range; In the fluorescence detection unit, the sample is excited based on the excitation wavelength set by the reference model R, and its corresponding fluorescence emission peak and intensity are recorded. In the electrochemical detection unit, by setting a linear scanning voltage range, the current response curve of the sample is recorded in real time, and the redox peak potential and peak current are extracted to construct a comprehensive detection signal S′.

[0011] Preferably, the output matching score P includes: The detection signal S′ is dimensionally split, and spectral, fluorescence, and electrochemical response sub-vectors are extracted respectively. Calculate the similarity between each response subvector and the feature vector of the corresponding additive in fingerprint feature group T; Calculate the set of similarity scores between all target additives and S′, and determine that the corresponding additive exists in the sample when the matching score is higher than a preset threshold; Output the matching results and the matching score P.

[0012] Preferably, the anomalous features are extracted, and a feature vector of the component to be verified is constructed based on the anomalous features, including: After determining that the matching score of all target additives is lower than the confidence threshold, abnormal pattern recognition is performed on the spectral, fluorescence and electrochemical subchannel data in the detection signal S′, and the signal segment deviating from the fingerprint feature group T is extracted by the local anomaly factor algorithm. Extract dominant features from the identified abnormal responses, including abnormal peak positions, amplitudes, and response timing parameters; The extracted multi-channel anomaly features are fused into a unified multi-dimensional vector format to construct the feature vector of the component to be verified.

[0013] Preferably, comparing the feature vector of the component to be verified with the template data of high-risk components includes: Retrieve a set of high-risk ingredient templates M from the illegal additives database. Each template contains a typical response feature vector of the ingredient under different food matrices. Feature vector of the component to be verified With each template vector Perform multidimensional similarity calculation; Set risk assessment threshold ,when With a certain template The similarity score is greater than At that time, it was determined that the component to be verified was associated with the corresponding high-risk component; Output the matching results and the templates that meet the conditions. It is marked as a suspected illegally added ingredient, and risk label information is generated at the same time.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a multi-channel response detection system driven by food matrix information, combining feature fusion and matching analysis of spectral, fluorescence, and electrochemical signals to effectively improve the breadth, real-time performance, and intelligent identification capabilities of food additive detection. In particular, by introducing additive fingerprint feature group T and a matching degree scoring mechanism, the system can achieve high-confidence rapid judgment in complex food matrix environments, overcoming the technical bottlenecks of traditional detection methods that rely on a single signal channel, have long detection cycles, and cannot handle unknown added components.

[0015] 2. This invention innovatively introduces an anomaly identification and illegal ingredient comparison module, which can identify potential illegal additive behavior when preset additives cannot be matched. By comparing the feature vector of the component to be verified with high-risk template data, it achieves intelligent early warning and preliminary classification identification of unknown or unregistered illegal additives. This technical solution has significant openness, scalability, and application adaptability, and is suitable for mobile rapid testing, on-site law enforcement, and risk screening in various food scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0019] For examples, please refer to Figure 1 As shown in this embodiment, a rapid detection system for food additive components includes: Sample collection module: Collects food samples to be tested and preprocesses them to obtain food matrix information F for analysis, where F includes the sample's pH value, viscosity, conductivity and color distribution parameters; Matrix modeling module: Based on the food matrix information F, it matches a preset food matrix model library to determine the reference model R that is closest to F; Additive feature extraction module: Based on the reference model R, extract the corresponding additive fingerprint feature group T from the additive feature response database, where T includes the predefined target additive's spectral absorption peak, fluorescence response value and electrochemical characteristic curve in the matrix; Multi-channel detection module: performs real-time multi-channel detection on the food sample and acquires the corresponding detection signal S′; Feature matching module: Performs feature matching analysis between the detection signal S′ and the fingerprint feature group T of the additive to obtain the judgment result of whether the target additive exists in the current sample, and outputs the matching score P; Anomaly detection module: When the matching score P is less than the preset confidence threshold, it is determined that there is an unknown additive in the sample, its abnormal features are extracted, and a feature vector of the component to be verified is constructed based on the abnormal features; Illegal ingredient identification module: compares the feature vector of the ingredient to be verified with the template data of high-risk ingredients, and marks suspected illegally added ingredients.

[0020] The sample acquisition module of this invention is used for the initial acquisition and preprocessing of food samples to be tested, aiming to provide stable and quantifiable matrix information support for subsequent rapid detection. This module includes a food sample acquisition unit and a sample preprocessing unit.

[0021] In practical applications, solid, liquid, or semi-fluid food samples can be collected directly from markets, restaurants, production lines, or unknown sources, depending on the testing requirements. To ensure sample representativeness and subsequent testing accuracy, sterile sampling tools are preferred, and sampling and transfer should be completed quickly to avoid contamination or alteration of the physicochemical properties of the samples due to environmental factors.

[0022] After sample collection is completed, the sample pretreatment process begins. This pretreatment includes, but is not limited to, mechanical stirring, dilution, filtration, and homogenization. If necessary, cooling, heating, or pH adjustment may be performed to meet the technical requirements of different detection channels.

[0023] After preprocessing, the sample will enter the food matrix parameter acquisition process, which involves extracting and detecting basic physicochemical indicators related to the sample using an integrated multi-parameter sensing component. Specifically, the food matrix information F acquired by this invention includes the following parameters: pH value: reflects the acidity or alkalinity of a food sample and affects the ionization degree and response signal of some additives; Viscosity: Used to determine the flow characteristics of a sample, affecting the stability of sample transport in the detection channel; Electrical conductivity: characterizes the concentration of dissolved ions in a sample, and indirectly reflects the presence and concentration of electrically active components in the additive; Color distribution parameters: The visible light absorption or reflection characteristics of the sample are analyzed by optical sensing components to provide a comparison benchmark for subsequent spectral matching.

[0024] The matrix modeling module in this invention is used to match the food matrix information F of a sample with a pre-set food matrix model library to determine the reference model R that is closest to the characteristics of the sample. This module enables rapid classification of samples with different food types, physicochemical properties, and testing environments, providing highly adaptable model support for subsequent additive feature extraction and response analysis. The matrix modeling module mainly includes the following processing steps: First, the original parameter values ​​in the food matrix information F obtained by the sample collection module are standardized to eliminate the interference caused by the differences in numerical magnitude between different parameters in subsequent similarity calculations.

[0025] The food matrix information F includes, but is not limited to, pH value, viscosity, conductivity, and color distribution parameters. These are denoted as follows: To achieve a unified dimension, the Z-score standardization method is adopted for each parameter. Calculate its standardized value The specific formula is: subtract the mean of each parameter from the historical samples, and then divide by the standard deviation. ,in, For parameter f i The mean in the baseline sample, This corresponds to the standard deviation. Finally, a food matrix feature vector under a unified dimension is constructed. .

[0026] Based on the constructed food matrix feature vector F′, its similarity is calculated with the feature vector of each preset reference model in the food matrix model library. The food matrix model library is an offline database containing matrix feature models for several typical food types. Each model Feature vectors with the same dimension as F′ To ensure matching accuracy and algorithm flexibility, this invention provides two similarity calculation methods: Euclidean distance algorithm: Calculate F′ and The distance between two entities in multidimensional space; the smaller the distance, the higher the similarity. Cosine similarity algorithm: Calculate F′ and... The similarity is determined by the cosine of the angle between the two points; the closer the cosine is to 1, the higher the similarity. The algorithm is automatically selected based on the actual detection scenario, or a matching strategy can be preset by the user.

[0027] To avoid false matches, this invention sets a similarity filtering threshold. The specific value can be set to 0.85 (when using cosine similarity) or the maximum Euclidean distance can not exceed D, depending on the application requirements. t (e.g., 3.0). Combine F′ with all... The calculated similarity result is compared with the set threshold. The comparison is performed to select a set of reference models that meet the conditions, and the model with the highest similarity to F′ is selected as the final matching reference model R.

[0028] The additive feature extraction module in this invention is used to extract a set of additive fingerprint features T that are compatible with a pre-set additive feature response database, based on a matched food matrix reference model R. The core objective of this module is to extract additive response features that still have good recognition capabilities under the current matrix conditions, based on the differences in the food matrix environment, for subsequent detection signal comparison and result determination.

[0029] First, based on the reference model R output by the matrix modeling module, the corresponding food matrix category information is extracted from it. This matrix category information is used to limit the search scope of the additive characteristic response database, avoiding distortion or misjudgment of response characteristics due to differences in the physicochemical environment of different food types.

[0030] Specifically, the additive characteristic response database is indexed by food matrix category, including more than ten common food types such as solid dairy products, fruit and vegetable beverages, high-salt pickled foods, and high-fat fried foods. The system uses... To retrieve tags, a set of target additives that are compatible with this type of matrix environment is selected from the database. Each additive The database contains complete records of multi-source response data.

[0031] For each additive in the selected target additive set A, multi-source response data was extracted for it under the specific food matrix environment. The multi-source response data included three detection dimensions: Spectral absorption peak information ( ): In the absorption curves of the visible or ultraviolet regions, record the position of the main absorption wavelength, peak height and peak width.

[0032] Fluorescence response value (Sb): The emission wavelength and peak fluorescence intensity recorded at a preset excitation wavelength; Electrochemical potential change curve (Sc): parameters such as redox potential and peak current recorded by cyclic voltammetry or differential pulse voltammetry.

[0033] The three types of response data mentioned above come from different detection platforms (such as spectrometers, fluorescence detectors, and electrochemical workstations), and their original dimensions and units are inconsistent. To achieve cross-channel fusion analysis, this invention uses a response behavior extraction method under a unified timestamp for Sa, Sb, and Sc, and encodes them into equal-dimensional vectors, which are then concatenated to form a fusion response vector Vi = [Va, Vb, Vc] to describe additive A. i A complete picture of the response characteristics under the current matrix conditions.

[0034] After multi-source data fusion, the vector Vi is further subjected to dimensionality reduction. The principal component analysis (PCA) algorithm used follows these steps: Normalization: Zero-mean and unit-variance transformations are applied to all components in Vi to eliminate numerical bias. Covariance Matrix Construction: The covariance matrix C of all fused vectors is constructed. Eigenvalue Decomposition: Eigenvalue and eigenvector decompositions are performed on the covariance matrix C. Principal Component Selection: The top K principal components with a cumulative contribution rate of over 90% are selected to construct dimensionality-reduced vectors. Dimensionally reduced feature vectors It retains the most discriminative response dimension, which can effectively characterize the key response behavior of additives in the current matrix environment.

[0035] To achieve model generalization and rapid comparison, the dimensionality-reduced feature vectors are further refined. Standardized coding is performed. Min-Max normalization is used to linearly map the values ​​of each dimension to the [0, 1] interval. Specifically, for a given dimension value x, it is converted to... ,in and These are the minimum and maximum values ​​of this dimension in the historical samples.

[0036] The processed feature vectors are stored as standard fingerprint feature sets. n is the total number of fingerprint features; where each Corresponding to additives And establish a one-to-one correspondence with the reference model R.

[0037] The multi-channel detection module in this invention is used to acquire signals from potential additives in the food sample in real time through multiple parallel detection channels after the food sample has been preprocessed and matched with the reference model R, and outputs a structured detection signal S′. This module achieves the coordinated acquisition of three types of signals: spectral detection, fluorescence detection, and electrochemical detection, ensuring complete correspondence with the response dimension of the aforementioned fingerprint feature group T, thereby providing a data foundation for subsequent feature matching analysis. The multi-channel detection operation includes the following steps: First, the pretreated food sample is divided into three equal parts according to a set ratio, and then injected into the spectral detection unit, fluorescence detection unit, and electrochemical detection unit, respectively. Each detection unit is an independent module with its own sensor interface, signal processing chip, and data transmission channel, and can perform the detection process simultaneously or sequentially.

[0038] To ensure consistent detection and aligned responses, this invention preferably employs an integrated control chip (such as an ARM Cortex-M series controller) with multi-channel synchronous control capabilities to uniformly regulate the sampling time, sampling frequency, and temperature control conditions of the three detection channels.

[0039] In the spectral detection unit, a broadband LED or xenon lamp is used as the excitation light source to illuminate the food sample through an integrated optical path system. The absorption and transmission behavior of the sample to light of a specific wavelength is processed by a spectrometer, and the full-band signal is acquired in real time by a photoelectric sensor array.

[0040] The acquired raw data constitute a wavelength-absorbance two-dimensional matrix, denoted as... Where λ represents the spectral wavelength in nanometers (nm), the data range typically covers 200 nm to 800 nm. The system automatically normalizes the light intensity of the curve and records parameters such as the wavelength position, peak height, and full width at half maximum (FWHM) of the main absorption peaks. The acquisition frequency is generally set between 10 Hz and 50 Hz, adjusted according to the sample transmittance and noise level. Finally, the processed absorption curve is encoded into a structured feature vector. This is used to compare the spectral absorption characteristics with those in T.

[0041] In the fluorescence detection unit, a specific excitation wavelength is selected based on the matrix type and additive type defined by the reference model R. The sample is irradiated using a laser or LED excitation source. The detection module is equipped with a high-sensitivity photomultiplier tube (PMT) or CMOS fluorescence detector to acquire the fluorescence emission signal generated by the sample in real time and generate emission spectral data within a specified time window. ,in To excite the wavelength, The emission wavelength is used. The wavelength and peak height of the fluorescence peak in the spectral line serve as core response features, forming a fluorescence response vector. .

[0042] In the electrochemical detection unit, the sample is injected into a microfluidic chip or electrochemical cell, and a control voltage is applied to it through an embedded electrode assembly (including a working electrode, a reference electrode, and an auxiliary electrode).

[0043] The detection process employs electrochemical techniques such as cyclic voltammetry (CV) or differential pulse voltammetry (DPV) at a specific scan rate (e.g., 50 mV / s) and voltage range (e.g., ...). Within +1.0 V, the current response of the sample is recorded to form a potential-current curve Sv(V,I).

[0044] Automatically identify peak current position, potential value, and peak area; extract parameters of main reduction and oxidation peaks; and construct electrochemical response vector. Meanwhile, to avoid signal drift, the system sets a baseline correction algorithm and sets a matching tolerance ΔV (e.g., ±0.05 V) based on the peak distribution of historical samples.

[0045] After acquiring the signals from each detection channel, the spectral response vectors are... The fluorescence response vector Sb′ and the electrochemical response vector Sc′ are merged according to predefined dimensions to construct a detection signal S′ with a unified structure: All sub-vectors have been normalized to ensure consistent signal amplitude and dimension, facilitating subsequent item-by-item matching analysis with the additive fingerprint feature group T.

[0046] The feature matching and judgment module in this invention is used to perform item-by-item matching analysis between the detection signal S′ obtained after multi-channel detection of food samples and the previously extracted additive fingerprint feature group T. Through quantitative similarity calculation, the module outputs the judgment result of whether each target additive exists in the current sample and the matching score P.

[0047] First, a structured detection signal S′ is received, which consists of three sub-channel data: spectral response vector Sa′, fluorescence response vector Sb′, and electrochemical response vector Sc′.

[0048] To achieve targeted matching, the system performs dimensional splitting of S′ and sends each sub-vector to the corresponding matching engine for corresponding calculation with various response features in T.

[0049] To evaluate the matching degree between the current sample signal and the fingerprint features of each target additive, this invention introduces a weighted cosine similarity algorithm to compare each feature vector in S′ and T one by one. The specific implementation is as follows: Vector standardization: First, the corresponding feature vectors in Sa′, Sb′, Sc′, and T are normalized using the L2 norm to ensure that the similarity calculation only reflects directional differences and avoids amplitude influence. Considering the different influences of different food matrices on the response of each channel, the system sets the channel weights W = {wa, wb, wc} based on the reference model R, where: wa is the weight of the spectral channel; wb is the weight of the fluorescence channel; wc is the weight of the electrochemical channel; typically, wa + wb + wc = 1. This weight can be determined through regression modeling based on the recognition contribution rate of each channel in historical training samples. Calculating channel similarity: The cosine similarity between the detection signal and the target additive in the three channels is calculated separately, expressed as: ; ; ; Calculation of overall matching score P: The score P∈[0, 1] indicates that the higher the value, the closer the current sample signal is to the target additive features.

[0050] To achieve highly reliable additive identification, this invention introduces a matching degree threshold. As a criterion. The threshold. Based on the statistical results of the matching score distribution in the training sample set under conditions of actual presence and absence of additives, the optimal range is set between 0.80 and 0.90, which can be flexibly adjusted according to different types of additives. The judgment rule is as follows: When : Determine if the target additive is present in the current food sample; when The algorithm determines whether the additive was not detected or was below the detection limit in the sample. This determination logic can be flexibly switched to an interval confidence mechanism based on sample complexity and regulatory requirements, i.e., setting a suspicious interval (e.g., 0.75 ≤ P < 0.80) and marking it as an intermediate result requiring further confirmation. The matching result for each target additive is encapsulated as a structured data entry, in the following format: [Additive Number] Match score Judgment result ];in This is a Boolean flag indicating whether the item exists.

[0051] In a rapid detection method for food additive components provided by this invention, an anomaly identification module is used to further identify potentially unknown additives in a sample when no fingerprint features can be matched with any known additives. This module extracts atypical response features by performing anomaly pattern recognition on the detection signal S′, and constructs a preliminary response feature vector for the component to be verified.

[0052] When the matching score P of all target additives is lower than the preset confidence threshold, the anomaly identification process is automatically triggered. The confidence threshold indicates the credible boundary of the similarity between the current detection signal and any fingerprint feature group T in the preset database. It is usually obtained through large-scale training sample statistics and is preferably set between 0.85 and 0.90.

[0053] If the matching score of all target additives is less than a preset confidence threshold, the current sample is determined to potentially contain unknown additives, and the anomaly detection process is initiated. To identify atypical response fragments that may be contained in the detection signal S′, the system uses the Local Outlier Factor (LOF) algorithm to detect anomalies in each channel's data. LOF Algorithm Introduction and Adaptation: The Local Outlier Factor (LOF) is a density-based anomaly detection method suitable for high-dimensional detection data with nonlinear and non-normal distributions. Its core idea is: if the local density of a data point is significantly lower than that of its neighboring data points, then that point is considered a local anomaly. Implementation Steps: The detection signal S′ is split into spectral signal Sa′, fluorescence signal Sb′, and electrochemical signal Sc′; the response vector of each channel is segmented using a sliding window (e.g., 20 sampling points per group) to construct a response fragment sequence; the density distribution of each fragment is calculated and compared with the densities of the K nearest neighbor fragments (e.g., K=5) to obtain the Local Outlier Factor (LOF) value; an anomaly judgment threshold is set (e.g., LOF>1.5), and all response fragments exceeding the threshold are marked as anomalous segments. Output an abnormal response index set for each channel. These fragments contain signal features that are inconsistent with known fingerprint responses, serving as candidate data regions for subsequent feature extraction.

[0054] For the identified abnormal response segments, the system further extracts representative dominant features, mainly including: peak position, amplitude, and response timing parameters.

[0055] For anomalous response regions, the first derivative method or Gaussian fitting method is used to identify the main absorption peak, fluorescence emission peak, or electrochemical polarization potential. Peak positions are expressed as wavelength (nm), potential (V), or time point (s).

[0056] Record the maximum amplitude value (such as absorbance value, fluorescence intensity, current peak value) corresponding to each main peak as a key parameter reflecting the response intensity.

[0057] Extract dynamic parameters such as the relative time position of the peak in the response curve, rise / fall time, and peak width to form a behavioral description of the response process.

[0058] The extracted features can be structured into the following form: F={ Main peak location Amplitude Time point, Peak width, : slope}, each channel independently extracts a set of features.

[0059] To achieve joint feature representation across channels, a unified vector fusion of spectral, fluorescence, and electrochemical anomaly features is performed to construct the response feature vector of the component to be verified. .

[0060] Align the dominant anomaly features of different channels to a unified vector format, set a fixed dimension (e.g., extract 5 features per channel, for a total of 15 dimensions), and pad with zeros or perform linear interpolation if necessary.

[0061] The Min-Max normalization method is used to normalize all features to the [0,1] interval to ensure that different physical quantities are comparable.

[0062] The features from each channel are concatenated to construct the final feature vector of the component to be verified. ,like: ,in Represents the i-th spectral feature. Represents the i-th fluorescence feature. This represents the i-th electrochemical feature.

[0063] The illegal component identification module provided by this invention is used to identify unknown abnormal responses during the detection process and then convert the constructed feature vector of the component to be verified. The system compares the samples against a pre-set set of high-risk ingredient templates in an expanded database of illegal additives to identify potential illegal additives and then labels and reports them based on their risk level.

[0064] First, a set of high-risk ingredient templates was extracted from an expanded database of illegal additives. This template set is built based on historical case samples, regulatory notification data, and frequently detected abnormal components.

[0065] Each template Each contains standard response features under multiple matrix environments, uniformly encoded into a structured vector format. It includes the following data elements: template number and substance name (e.g., dibutyl phthalate, sodium formaldehyde sulfoxylate, etc.); response feature vector: Dimensions and vectors to be verified Consistency; Adaptive matrix labeling: used to record the type of food matrix in which the template was constructed; Risk level labeling (e.g., Category I prohibited, Category II restricted). The database supports incremental updates and can be continuously supplemented through expert verification, laboratory retesting, or synchronization with regulatory data.

[0066] The feature vector of the component to be verified generated in the previous module With each template vector in the template set M A comparison is performed. Since features from different channels may exhibit statistical correlation, this module uses Mahalanobis distance as the similarity calculation method. The Mahalanobis distance algorithm is defined as follows: Let the vector to be verified be... The template vector is Both are vectors with the same dimension; Calculate Mahalanobis distance The expression is: ;in: Let S be the difference vector; S is the covariance matrix of all template features. Let S be the inverse matrix of S; T denotes the transpose operation.

[0067] Since a smaller Mahalanobis distance indicates a higher similarity between two vectors, the final similarity score is calculated using its reciprocal or inverse function form. ,like: ; To avoid misjudgment, this invention introduces a risk assessment threshold. This threshold is used to determine whether the comparison result can be classified as a high-risk component. This threshold can be set through empirical statistics on the training set or by an expert system, with a preferred range of 0.80 to 0.90. The comparison logic is as follows: If a certain template Similarity score If the component to be verified is highly correlated with the template, it is identified as a suspected illegally added component. If all If so, it is determined that there are unknown abnormal added components in the sample that have not yet been modeled, and it is recommended to include them in the list of items to be built.

[0068] The comparison results will be output as structured results, including the following: the name of the suspected illegal component (e.g., template). (Referring to the components); matching score Risk level (Class I / Class II, etc.); Matching source description (e.g., with the vector to be verified) (Which features have a high matching degree); labeling time and detection batch information.

[0069] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A rapid detection system for food additive components, characterized in that: include: Sample collection module: Collects food samples to be tested and preprocesses them to obtain food matrix information F for analysis, where F includes the sample's pH value, viscosity, conductivity and color distribution parameters; Matrix modeling module: Based on the food matrix information F, it matches a preset food matrix model library to determine the reference model R that is closest to F; Additive feature extraction module: Based on the reference model R, extract the corresponding additive fingerprint feature group T from the additive feature response database, where T includes the predefined target additive's spectral absorption peak, fluorescence response value and electrochemical characteristic curve in the matrix; Multi-channel detection module: performs real-time multi-channel detection on the food sample and acquires the corresponding detection signal S′; Feature matching module: Performs feature matching analysis between the detection signal S′ and the fingerprint feature group T of the additive to obtain the judgment result of whether the target additive exists in the current sample, and outputs the matching score P; Anomaly detection module: When the matching score P is less than the preset confidence threshold, it is determined that there is an unknown additive in the sample, its abnormal features are extracted, and a feature vector of the component to be verified is constructed based on the abnormal features; Illegal ingredient identification module: compares the feature vector of the ingredient to be verified with the template data of high-risk ingredients, and marks suspected illegally added ingredients.

2. The rapid detection system for food additive components according to claim 1, characterized in that: The step of matching the food matrix information F with a preset food matrix model library to determine the reference model R that is closest to F includes: The parameters in the acquired food matrix information F are standardized to construct a matrix feature vector F′ under a unified dimension. The similarity between F′ and the feature vectors of each reference model in the preset food matrix model library is calculated using the Euclidean distance algorithm or the cosine similarity algorithm. Set a similarity threshold, and select the reference model R with the highest similarity that is greater than the threshold based on the calculation results.

3. The rapid detection system for food additive components according to claim 2, characterized in that: Based on the reference model R, the corresponding additive fingerprint feature group T is extracted from the additive feature response database, including: The food matrix category information corresponding to the reference model R is input into the additive feature response database to limit the search scope and filter the target additive set that matches the matrix environment. Feature fusion is performed on the multi-source response data of each additive in the selected target additive set. The multi-source response data includes spectral absorption peaks, fluorescence response values, and electrochemical potential change curves. A dimensionality reduction algorithm based on principal component analysis was used to extract features from the fused multidimensional response data to obtain feature vectors characterizing the key response behaviors of each additive. The feature vectors are standardized and stored as additive fingerprint feature groups T, and a one-to-one correspondence is established with the reference model R.

4. The rapid detection system for food additive components according to claim 3, characterized in that: Where T includes the predefined target additive's spectral absorption peaks, fluorescence response values, and electrochemical characteristic curves in the matrix, including: The matrix parameters corresponding to the reference model R are called, and the response regulation factor of the additive in this type of matrix environment is set to correct the baseline state of the additive response characteristics. Based on the regulatory factors, the original response data of each target additive in the database were adapted and transformed to obtain the response reconstruction results of the spectral absorption curve, fluorescence emission intensity curve and electrochemical potential change curve under this type of matrix. The main absorption peak position in the spectral curve is identified by the peak extraction algorithm, the maximum excitation value and corresponding emission value of the fluorescence response are extracted, and the main reduction / oxidation potential range in the electrochemical response is extracted to form a standard response feature subset.

5. The rapid detection system for food additive components according to claim 1, characterized in that: Obtain the corresponding detection signal S′, including: The pretreated food samples were assigned to the spectral detection unit, fluorescence detection unit, and electrochemical detection unit, respectively, corresponding to various response dimensions in the additive fingerprint feature group T. In the spectral detection unit, a broadband light source is used to illuminate the sample and collect the reflected or transmitted signals to generate a spectral absorption curve within the wavelength range; In the fluorescence detection unit, the sample is excited based on the excitation wavelength set by the reference model R, and its corresponding fluorescence emission peak and intensity are recorded. In the electrochemical detection unit, by setting a linear scanning voltage range, the current response curve of the sample is recorded in real time, and the redox peak potential and peak current are extracted to construct a comprehensive detection signal S′.

6. The rapid detection system for food additive components according to claim 5, characterized in that: The output matching score P includes: The detection signal S′ is dimensionally split, and spectral, fluorescence, and electrochemical response sub-vectors are extracted respectively. Calculate the similarity between each response subvector and the feature vector of the corresponding additive in fingerprint feature group T; Calculate the set of similarity scores between all target additives and S′, and determine that the corresponding additive exists in the sample when the matching score is higher than a preset threshold; Output the matching results and the matching score P.

7. The rapid detection system for food additive components according to claim 6, characterized in that: Extracting its abnormal features, And construct the feature vector of the component to be verified based on the abnormal features, including: After determining that the matching score of all target additives is lower than the confidence threshold, abnormal pattern recognition is performed on the spectral, fluorescence and electrochemical subchannel data in the detection signal S′, and the signal segment deviating from the fingerprint feature group T is extracted by the local anomaly factor algorithm. Extract dominant features from the identified abnormal responses, including abnormal peak positions, amplitudes, and response timing parameters; The extracted multi-channel anomaly features are fused into a unified multi-dimensional vector format to construct the feature vector of the component to be verified.

8. The rapid detection system for food additive components according to claim 7, characterized in that: The comparison of the feature vector of the component to be verified with the template data of high-risk components includes: Retrieve a set of high-risk ingredient templates M from the illegal additives database. Each template contains a typical response feature vector of the ingredient under different food matrices. Feature vector of the component to be verified With each template vector Perform multidimensional similarity calculation; Set risk assessment threshold ,when With a certain template The similarity score is greater than At that time, it was determined that the component to be verified was associated with the corresponding high-risk component; Output the matching results and the templates that meet the conditions. It is marked as a suspected illegally added ingredient, and risk label information is generated at the same time.

Citation Information

Cited By

  • Sulfur dioxide detection system for canned food

    CN122193359A