Intelligent detection method for quantifying linseed health product additives based on spectral analysis

Abstract

The application discloses a flaxseed health product additive quantitative intelligent detection method based on spectral analysis. The method adopts a double-channel integrated spectrometer to non-destructively in-situ collect mid-infrared and near-infrared spectra of flaxseed health product finished products, uses a physical information constraint deep residual network to embed a Kubelka-Munk scattering model and a temperature drift model in the form of a constraint loss item for training, and performs self-adaptive baseline correction on the original spectrum. Then, a pre-constructed auxiliary material spectrum dictionary is called, L1 regularization sparse coding is used to deduct the auxiliary material contribution spectrum, and multivariate curve resolution-alternating least squares iterative decomposition is used to obtain pure flax lignan SDG and pure cyanogenic glycoside component spectrum. Finally, a bidirectional orthogonal partial least squares method is used for joint variation modeling of the double-channel pure spectrum, and quantitative prediction values of the SDG content and the cyanogenic glycoside content are synchronously output.

Description

Technical Field

[0001] This invention belongs to the field of food safety testing technology, specifically relating to a quantitative intelligent detection method for flaxseed health product additives based on spectral analysis. Background Technology

[0002] Flaxseed is rich in functional lignans, the main active form of which is open-ring isolar resinol disodium glycoside, comprising approximately 0.3%–1.0% of its dry weight. This is a core efficacy indicator component in flaxseed health products. Simultaneously, flaxseed naturally occurring cyanogenic glycosides, primarily linamarin and linalool, can release hydrocyanic acid after being metabolized in the human body, exhibiting a certain degree of toxicity and falling under mandatory safety testing. With the rapid expansion of the market for health products using these components as core efficacy additives, product quality control faces severe challenges.

[0003] The current mainstream detection method is high-performance liquid chromatography (HPLC). This method requires grinding the sample and ultrasonically extracting it with an organic solvent. After purification by solid-phase extraction, the sample is injected and gradient eluted using a reversed-phase C18 column to obtain the chromatographic peaks of SDG and cyanogenic glycosides for quantitative calculation. This method has the following drawbacks: First, it is a destructive detection method, and the sample cannot be recovered, making it unsuitable for large-scale online sampling in production lines; second, SDG and cyanogenic glycosides usually require separate methods and multiple injections, making it impossible to obtain the content of both simultaneously in a single test; third, a single test takes more than 60 minutes, consumes a large amount of organic solvent, generates waste liquid containing organic matter, and has high detection costs.

[0004] The existing scheme two is a single near-infrared partial least squares quantitative model. This scheme acquires the near-infrared spectrum of the sample, performs standard normal variable transformation preprocessing, and then uses partial least squares regression modeling with the spectral matrix and HPLC reference values ​​to predict new samples online. This scheme has the following drawbacks: First, both SDG and cyanogenic glycosides contain glycosidic bonds in their molecular skeletons, and their absorption peaks overlap significantly in the 2050–2100 nm range of the near-infrared spectrum. Single-band modeling cannot distinguish between them, and the residual amount of cyanogenic glycosides cannot be effectively detected. Second, the finished product contains various excipients such as microcrystalline cellulose, magnesium stearate, and gelatin, whose strong spectral absorptions severely overlap with the characteristic absorption bands of the target components, making the PLS linear model prone to overfitting or prediction failure. Third, inter-batch particle size distribution and environmental temperature fluctuations cause spectral baseline drift. Traditional preprocessing methods have limited effectiveness in complex solid-liquid mixed systems, resulting in weak model generalization ability and frequent remodeling, failing to meet the industrial requirements for long-term stable online detection.

[0005] In summary, existing technologies cannot simultaneously achieve non-destructive, synchronous, and rapid quantitative detection of SDG and cyanogenic glycosides in flaxseed health products, and there is an urgent need for a new detection method that can overcome the above-mentioned shortcomings. Summary of the Invention

[0006] The technical problem to be solved by this invention is to overcome the three core defects in the prior art: the inability to quantify SDG and cyanogenic glycosides synchronously due to near-infrared spectral overlap, the strong spectral interference caused by complex excipient matrix in finished dosage forms, and the poor long-term stability of the model due to fluctuations in operating conditions. This invention provides a non-destructive, synchronous, and rapid intelligent method for the quantitative detection of flaxseed health product additives.

[0007] This invention provides a quantitative intelligent detection method for flaxseed health product additives based on spectral analysis, which is performed according to the following steps: Step 1: Dual-channel spectral acquisition A dual-channel integrated spectrometer equipped with an attenuated total reflectance accessory was used to perform non-destructive in-situ spectral acquisition of the flaxseed health product under test, without the need to open or grind the sample. The mid-infrared channel acquisition wavenumber range was 650–4000 cm⁻¹. - ¹, Resolution is 4 cm - The spectrum of the first channel was scanned a total of 32 times; the near-infrared channel acquired spectra with a wavelength range of 1000–2500 nm and a resolution of 1 nm, also scanned a total of 32 times. The raw spectral data from both channels were converted by absorbance to form mid-infrared spectral matrices. With near-infrared spectral matrix ,in For the number of samples, and These are the wavenumber variables for the mid-infrared and near-infrared regions, respectively.

[0008] Step 2: Physical Information Constraint Adaptive Baseline Correction Before the spectrum enters the signal analysis step, the original spectrum is first adaptively baseline corrected by a Physical Information Constrained Deep Residual Network (PINN-ResNet) to eliminate light scattering interference caused by particle size changes and the influence of ambient temperature drift on the baseline.

[0009] Based on the Kubelka-Munk light scattering theory, the observed absorbance of solid samples under diffuse reflection conditions satisfy:

[0010] in, The true absorbance is determined solely by the chemical composition; This is a wavelength-dependent scattering coefficient function; Median particle size of the current sample volume (unit: μm); This is the median reference particle size of the sample group during modeling; Temperature drift coefficient (unit: AU / °C); The current measured temperature (unit: °C); The reference temperature for modeling is usually 25°C.

[0011] PINN-ResNet consists of two branches: a spectral feature extraction backbone (composed of four stacked residual blocks, each containing a one-dimensional convolutional layer, a batch normalization layer, ReLU activation, and residual connections) and a physical constraint prior layer (two fully connected layers, with granular parameters). and measuring temperature Using Softplus as the activation function as input, the output is a scattering coefficient vector. and temperature drift coefficient scalar The intermediate features of the two branches are fused through Hadamard product, and the network outputs a wavelength-wise dynamic multiplicative correction coefficient vector. Additive bias vector The corrected spectrum is:

[0012] in, This is the original spectral vector; Represents element-wise product; This is the corrected spectral vector.

[0013] The total loss function for network training consists of two parts: prediction loss and physical constraint loss.

[0014]

[0015]

[0016] in, This is the weighting coefficient for the physical constraint loss; an initial value of 0.1 is recommended. For the first The model-predicted content vector for each sample; This is the corresponding HPLC reference content vector; This is the additive bias vector output by the network; The temperature drift coefficient output by the physical prior layer; This is the scattering coefficient vector output by the physical prior layer.

[0017] Step 3: Construction of the excipient spectral dictionary In the offline phase, at least eight kinds of pharmaceutical excipients commonly used in health product manufacturing (including microcrystalline cellulose, magnesium stearate, gelatin, croscarmellose sodium, fumed silica, hydroxypropyl methylcellulose, talc, and sucrose) were selected. Mid-infrared and near-infrared spectral vectors were collected for each excipient, with each excipient's data collected at least 10 times and averaged. The spectral vectors of each excipient were then normalized using the L2 norm.

[0018] in, For the first The original spectral vector of the auxiliary material; Its L2 norm; This represents the normalized spectral vector. All normalized excipient spectral vectors are concatenated column-wise to form mid-infrared excipient dictionary matrices. Near-infrared auxiliary material dictionary matrix ,in This represents the number of auxiliary material types. The dictionary is persistently stored in a structured database format.

[0019] Step 4: Extraction of Sparse Coding Auxiliary Material Contributions Using the auxiliary material dictionary matrix Based on, the corrected mixed spectral vector Solving for the sparse coding coefficient vector The objective function is optimized as follows:

[0020] in, For the optimal sparse coefficient vector, its th The element represents the first The contribution weight of each excipient in the mixed spectrum; The L1 regularization coefficient is recommended to have an initial value range of 0.001 to 0.01, which is determined through cross-validation. The L1 norm ensures that the solutions are sparsity; constraints To ensure physical interpretability, this optimization problem is solved iteratively using the coordinate descent method. The contribution spectra of the excipients are reconstructed using the obtained sparse coefficients. And extract the residual spectrum:

[0021] in, The residual spectral vector after removing the contribution of excipients mainly includes the spectral contribution of the active ingredients (SDG, cyanogenic glycosides).

[0022] Step 5: MCR-ALS in-situ blind unpacking The residual spectral vectors of multiple samples to be tested are stacked row by row to construct a residual spectral matrix. (The number of samples in this batch shall not be less than 20). The initial pure component spectral estimates are extracted using the Simplicity Detection Algorithm (SIMPLISMA), and the number of chemical components to be separated is set. These correspond to SDG, cyanogenic glycosides, linolenic acid, and other trace components, respectively. Based on a bilinear model... Under the premise of applying nonnegativity constraints and normalization constraints, the concentration matrix is ​​updated alternately. With pure component spectral matrix Iterate until convergence (the rate of change of the fitting residual between two adjacent iterations is less than...). (Or the number of iterations exceeds 500). The final output is the spectral vector of pure SDG components and the spectral vector of pure cyanogenic glycoside components, which are used as inputs for subsequent quantitative modeling.

[0023] Step Six: Quantitative Fusion of Two-Way Orthogonal Partial Least Squares Dual-Band Fusion The mid-infrared pure active ingredient separation spectrum matrix output from step five is denoted as... The near-infrared corresponding matrix is ​​denoted as After performing mean centering, the results are obtained. and Target response matrix The two columns represent the SDG content and cyanogenic glycoside content determined by the HPLC reference method, respectively, and were also mean-centered to obtain... .

[0024] Using bidirectional orthogonal partial least squares (O2PLS) and They are decomposed into joint variant components and their respective unique orthogonal variant components:

[0025]

[0026] in, for The joint score matrix; for The joint score matrix; The number of joint principal components is determined through cross-validation; and For their respective joint load matrices; , for Orthogonal at The score and load; , for orthogonal components; , Let each be a residual matrix. The joint component extraction criterion is to maximize... and The covariance between them is solved iteratively using the NIPALS algorithm.

[0027] Using the extracted joint score matrix Establish the target response matrix Ordinary least squares regression yields the regression coefficient matrix. :

[0028] in, The two columns correspond to the regression coefficient vectors of SDG and cyanogenic glycoside content, respectively. Five-fold cross-validation was used to evaluate the number of combined components. (Search range 1-15), Mid-infrared orthogonal components Orthogonal components of near-infrared (Each search range is 0-5) Perform a grid search and determine the optimal parameter combination with the goal of minimizing the sum of the root mean square errors of the cross-validation of SDG and cyanogenic glycoside predictions.

[0029] For new samples to be tested, the joint score vector is extracted based on the optimized O2PLS model parameters. Simultaneous prediction of SDG and cyanogenic glycoside content:

[0030] in, This is a vector of synchronized predicted values ​​for SDG and cyanogenic glycoside content in the new sample; It is a vector composed of the mean values ​​of each column of the response matrix of the training set.

[0031] Step 7: Output of Quantitative Results and Safety Assessment The predicted SDG content and cyanogenic glycoside content output in step six are compared with the safety threshold and the label claim range. If the predicted cyanogenic glycoside content exceeds the preset safety threshold, a safety warning is triggered and the batch is marked as unqualified. If the predicted SDG content deviates from the label claim range by more than 15%, the effective ingredient content is marked as substandard. Finally, a test report is generated that includes the SDG content, cyanogenic glycoside content, and the qualification judgment conclusion.

[0032] The beneficial effects of this invention are: 1. Utilizing the C≡N functional group of cyanogenic glycosides in the mid-infrared region at approximately 2200 cm⁻¹ - The strong characteristic stretching vibration absorption peak at ¹ was used to perform joint variation separation modeling with the near-infrared absorption characteristics of SDG using the O2PLS framework. This broke through the spectral overlap dead zone between the two in the near-infrared 2050–2100 nm range, and enabled the simultaneous and accurate quantification of toxic cyanogenic glycosides and functional SDG in a single scan, filling the technical gap that existing single-band spectral methods cannot distinguish between the two.

[0033] 2. By constructing an excipient spectral dictionary and combining sparse coding with a two-stage unwinding strategy of MCR-ALS, excipient interferences are mathematically separated from mixed spectra without damaging the integrity of the sample, restoring the pure spectral signal of the target active ingredient, thus realizing non-destructive in-situ detection of finished dosage forms.

[0034] 3. By embedding the Kubelka-Munk light scattering theory and temperature drift model into a deep residual network in the form of physical constraint loss terms, the correction behavior of the model conforms to physical laws, effectively compensating for the scattering coefficient variation caused by particle size change and the ambient temperature drift, and significantly extending the validity period of the quantitative model.

[0035] 4. The single detection cycle is shortened from more than 60 minutes for HPLC to no more than 30 seconds. No chemical reagents are required, realizing green online detection without reagents or waste liquid. It is suitable for large-scale, online, non-destructive quality inspection of health product production lines. Attached Figure Description

[0036] The invention will now be further described with reference to the accompanying drawings.

[0037] Figure 1 This is a diagram of the overall system architecture of the present invention, showing the composition and data flow of the spectral acquisition layer, signal analysis layer, and intelligent modeling layer; Figure 2 The flowchart for constructing the spectral dictionary for excipients and blind unrolling using MCR-ALS shows the detailed algorithm steps for the offline dictionary construction stage and the online blind unrolling stage. Figure 3 The flowchart for O2PLS dual-band fusion quantitative modeling shows the process of joint variant extraction, orthogonal variant filtering, and synchronous quantitative output of mid-infrared pure spectrum and near-infrared pure spectrum; Figure 4 The diagram shows the PINN-ResNet adaptive correction network structure, illustrating the network structure of the spectral feature extraction backbone, the physical constraint prior layer, the feature fusion module, and the corrected spectral output. Figure 5 This is a complete flowchart of the detection method of the present invention, showing the complete online detection steps and judgment branches from sample placement to output of the detection report. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] I. System Overall Architecture The present invention provides a quantitative intelligent detection method for flaxseed health product additives based on spectral analysis. The overall system architecture consists of three sequential functional layers, such as... Figure 1 As shown: The spectral acquisition layer contains two parallel acquisition channels, mid-infrared and near-infrared. The probe performs non-destructive in-situ spectral acquisition of the finished health product under test, and outputs the original MIR mixed spectral matrix and the original NIR mixed spectral matrix respectively. The signal analysis layer receives the original spectral matrix, which is then passed sequentially through the physical information constrained adaptive baseline correction network, the excipient spectral dictionary and the sparse coded excipient contribution subtraction module, and the MCR-ALS blind deconvolution module to output the pure SDG component spectrum and the pure cyanogenic glycoside component spectrum. The intelligent modeling layer receives the pure dual-channel spectrum output from the signal analysis layer, and through the O2PLS dual-band fusion quantitative model, simultaneously outputs two quantitative results: SDG content (mg / g) and cyanogenic glycoside content (mg / g).

[0040] II. Specific Implementation of Spectral Data Acquisition A dual-channel integrated spectrometer equipped with an attenuated total reflectance attachment was used to perform non-destructive in-situ scanning of the test samples. For hard capsule samples, the capsule was placed directly on the attenuated total reflectance probe plate without opening the shell, and constant pressure was applied before scanning. For compressed candies, the tablets were placed on the diffuse reflectance integrating sphere attachment for scanning.

[0041] Mid-infrared channel acquisition parameters: wavenumber range 650–4000 cm⁻¹ - ¹, resolution 4 cm - ¹, A total of 32 scans were performed to improve the signal-to-noise ratio; background spectra were collected once before each batch of samples was tested.

[0042] Near-infrared channel acquisition parameters: wavelength range 1000–2500 nm, resolution 1 nm, number of scans 32; background was acquired using a blank reference plate before each batch of tests.

[0043] The two-channel raw spectral data were converted by absorbance ( ,in The intensity of transmitted light from the sample. After setting the background reference light intensity, mid-infrared spectral matrices are formed respectively. With near-infrared spectral matrix ,in This refers to the quantity of samples in this batch. and These are the wavelength (wavenumber) variables for mid-infrared and near-infrared, respectively.

[0044] While acquiring dual-channel spectra, the median particle size of the current sample is determined by a laser particle size analyzer. The current measured temperature is recorded by the built-in temperature sensor. This serves as the physical parameter input to subsequent PINN-ResNet modules.

[0045] III. Construction of the Spectral Dictionary for Auxiliary Materials and Specific Implementation of In-situ Blind Decomposition of MCR-ALS like Figure 2 As shown, this step is divided into an offline dictionary construction stage and an online blind solution stage.

[0046] (I) Offline dictionary construction stage The following pharmaceutical excipients commonly used in health product manufacturing were selected, including but not limited to: microcrystalline cellulose, magnesium stearate, gelatin, croscarmellose sodium, fumed silica, hydroxypropyl methylcellulose, talc, and sucrose. Auxiliary materials, No fewer than 8 types.

[0047] Mid-infrared spectral vectors were collected for each type of excipient pure product. and near-infrared spectral vector (subscript) Representing the Each type of auxiliary material was sampled at least 10 times and the average was taken to reduce random measurement errors.

[0048] L2 norm normalization was performed on the spectral vectors of each excipient to eliminate the influence of excipient concentration differences on the vector magnitude.

[0049] in, This is the normalized spectral vector; for L2 norm, This represents the number of wavelength variables for the corresponding channel.

[0050] All normalized excipient spectral vectors are concatenated column-wise to form mid-infrared excipient dictionary matrices. Near-infrared auxiliary material dictionary matrix Each column corresponds to the normalized pure spectral vector of one excipient. The dictionary matrix is ​​persistently stored in the form of a structured database, recording the name, batch number, and collection time of the excipient corresponding to each column for subsequent updates and maintenance.

[0051] (II) Online Blind Decoding Stage – Extraction of Sparse Coding Auxiliary Material Contributions The sample spectral vector after step two PINN-ResNet correction As a mixed spectrum input. (Using the excipient dictionary matrix) (Mid-infrared channel corresponding) Near-infrared channel corresponding Using as a basis, solve for the sparse coding coefficient vector. :

[0052] In the above formula, The The element represents the first The contribution weight of each excipient in the mixed spectrum; The L1 regularization coefficient (recommended initial value range: 0.001–0.01, to be determined through cross-validation); L1 norm constraint. This ensures that the solution is sparsity to match the physical fact that the amount of excipients in actual formulations is finite; nonnegativity constraint Physical interpretability is guaranteed. This optimization problem is solved iteratively using the coordinate descent method, and it typically converges within 100 iterations.

[0053] Reconstructing the contribution spectrum of excipients using the obtained sparse coefficient vector And extract the residual spectrum. The residual spectrum mainly includes the spectral contributions of the active ingredient and other trace components, while the excipient signal has been significantly reduced.

[0054] (III) Online blind unpacking stage – MCR-ALS fine pure component separation The residual spectral vectors of all samples tested in this batch are stacked row-wise to construct a residual spectral matrix. (The number of samples in this batch shall not be less than 20). The SIMPLISM algorithm shall be used to analyze... The analysis automatically extracts the initial pure component spectral estimates. , of which components (Corresponding to SDG, cyanogenic glycosides, linolenic acid, and other trace components), corresponding to the initial concentration matrix. Obtained by least squares estimation.

[0055] Based on bilinear model (in For the concentration matrix, The spectral matrix of the pure components, (where the residual error matrix is ​​used), and iterative optimization is performed alternately under the conditions of applying non-negativity constraints and normalization constraints: fixed Least squares update :

[0056] After the update Set all negative values ​​to zero; fix Least squares update :

[0057] After the update Nonnegativity constraints and L2 norm normalization constraints are applied. The iterative convergence criterion is the rate of change of the fitting residuals between two consecutive iterations. satisfy ,in The iteration stops when the number of iterations exceeds 500, using the Frobenius norm (the square root of the sum of squares of all elements in the matrix) or the total number of iterations exceeds the standard norm. The final output is a pure SDG component spectral vector. ( (Row 1) and the spectral vector of pure cyanogenic glycoside components , The second row, and the corresponding concentration contribution matrix The first two columns.

[0058] IV. Specific Implementation of O2PLS Dual-Band Fusion Quantitative Modeling like Figure 3 As shown, the modeling phase uses a standard sample set containing known HPLC reference values, and it is recommended to train with no fewer than 100 samples.

[0059] The mid-infrared channel separation spectral matrix of pure active ingredients output by MCR-ALS is denoted as... The matrix corresponding to the near-infrared channel is denoted as Perform mean centering on each: ,in It is a column vector of all 1s. for A column vector formed by the mean values ​​of each column. Target response matrix. The two columns represent the HPLC reference values ​​for SDG content and cyanogenic glycoside content, respectively, and were also mean-centered to obtain the values. .

[0060] O2PLS will and It is decomposed into joint variation components and orthogonal variation components respectively. The extraction criterion for joint components is to maximize... and The covariance between them, through the analysis of The optimal joint load vector is extracted by eigenvalue decomposition and solved iteratively using the NIPALS algorithm.

[0061] Joint score matrix After extraction, establish a pair Ordinary least squares regression yields the regression coefficient matrix. The optimal number of joint components was determined using 5-fold cross-validation. The search range is 1–15, and the optimal orthogonal components are found in the mid-infrared and near-infrared regions. , Each has a search range of 0 to 5, with the optimization objective being to minimize the sum of the root mean square errors of the cross-validation of SDG and cyanogenic glycoside predictions.

[0062] During online prediction, steps two through five are performed sequentially on new samples to obtain pure MIR spectra. Compared with pure NIR spectrum After mean centering, the joint score vector is extracted using the optimized O2PLS model parameters. Substitute into the regression equation Simultaneously output the predicted values ​​of SDG content and cyanogenic glycoside content.

[0063] V. Specific Implementation of PINN-ResNet Training like Figure 4 As shown, each sample in the training dataset contains the original spectral vector. Particle size parameters measured by laser particle size analyzer Temperature measured by built-in sensor Reference content label verified by HPLC Through variable temperature and particle size experiments, with a temperature range of 20–35°C and a particle size range of 10–200 μm, the system collected spectral data of the same sample to construct a physical parameter-spectral drift control dataset for calculating physical constraint loss.

[0064] The spectral feature extraction backbone consists of four residual blocks cascaded together. Each residual block contains: a one-dimensional convolutional layer (64 filters, 11 kernel length) → a batch normalization layer → ReLU activation → a one-dimensional convolutional layer (64 filters, 11 kernel length) → a batch normalization layer → residual addition. The backbone output is then subjected to global average pooling to obtain the feature vector. The physical constraint prior layer is a two-layer fully connected network with 32 hidden layers, using Softplus activation, and receiving a vector of physical parameters. Output scattering coefficient vector and temperature drift coefficient scalar The intermediate features of the two branches are fused via Hadamard product, and the output is a vector of wavelength-wise dynamic multiplicative correction coefficients. Additive bias vector The corrected spectrum is .

[0065] The network uses the Adam optimizer (initial learning rate). , , The parameters are updated, and the training is performed for 200 epochs with a batch size of 32. An early stopping strategy is adopted (the training stops when the loss on the validation set no longer decreases after 20 consecutive epochs) to prevent overfitting. After training, the network weights are persistently saved as files and deployed in the embedded processing unit of the detection device.

[0066] VI. Specific Implementation of the Complete Online Testing Process like Figure 5As shown, the complete online detection process is as follows: The flaxseed health product sample to be tested is in place, no need to open or grind; dual-channel spectral acquisition is performed on the sample, and the current particle size is read simultaneously. With temperature Sensor data; the deployed PINN-ResNet is invoked to perform adaptive baseline correction on the raw spectrum, and the corrected spectrum is output. The excipient spectral dictionary is called, and sparse coding L1 regularization optimization is performed to reconstruct and subtract the contribution of excipients from the spectrum, thus obtaining the residual spectrum. Perform MCR-ALS iterative decomposition to output the pure SDG spectrum and the pure cyanogenic glycoside spectrum; call the trained O2PLS dual-band fusion model to simultaneously predict the SDG content. With cyanogenic glycoside content The system determines whether the predicted cyanogenic glycoside content exceeds a preset safety threshold. If it does, a safety warning is triggered, the batch is marked as substandard, and an alarm signal is sent. It also determines whether the predicted SDG content is within ±15% of the label's claimed range. If not, the active ingredient content is marked as substandard. A test report is generated, simultaneously outputting the SDG content, cyanogenic glycoside content, and pass / fail conclusion. From sample placement to obtaining the test report, the entire process takes no more than 30 seconds, requires no chemical reagents, and can be integrated into the online quality inspection node of a health supplement production line.

[0067] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A quantitative intelligent detection method for flaxseed health product additives based on spectral analysis, characterized in that, Includes the following steps: Non-destructive in-situ dual-channel spectral acquisition was performed on the flaxseed health product to be tested. The wavenumber range for the mid-infrared channel was 650–4000 cm⁻¹. -¹ The spectral matrix, with the near-infrared channel acquiring a spectral matrix in the wavelength range of 1000–2500 nm; A physically constrained deep residual network is used to perform adaptive baseline correction on the acquired raw spectra. The network incorporates scattering correction constraints and temperature drift correction constraints based on Kubelka-Munk light scattering theory as physical constraint loss terms during training. The network input includes the raw spectral vector and the simultaneously measured median particle size of the sample volume. and measuring temperature The network outputs a wavelength-wise dynamic multiplicative correction coefficient vector and an additive bias vector, which are used to perform element-wise multiplication and bias correction operations on the original spectrum to obtain the corrected spectrum. The pre-built excipient spectral dictionary is called, and the sparse coefficient vector of the corrected mixed spectrum relative to the excipient dictionary is solved by non-negative sparse coding with L1 regularization constraint. The excipient contribution spectrum is reconstructed and subtracted from the corrected mixed spectrum to obtain the residual spectrum. Multivariate curve-resolved alternating least squares iterative decomposition was performed on the residual spectral matrix, with nonnegativity and normalization constraints applied. The number of separated components was set to 3, corresponding to linolenic acid SDG, cyanogenic glycosides, linolenic acid, and other trace components, respectively. The pure SDG component spectrum and the pure cyanogenic glycoside component spectrum were output. The pure active ingredient spectral matrices output from the mid-infrared and near-infrared channels were mean-centered. The joint variation components carrying concentration information in both channels were extracted using a two-way orthogonal partial least squares method. At the same time, the unique orthogonal interference variables of each channel were filtered out. Based on the joint score matrix, a synchronous least squares regression was established for SDG content and cyanogenic glycoside content, and the quantitative predicted values ​​of SDG content and cyanogenic glycoside content in the sample were output synchronously. The predicted values ​​of SDG content and cyanogenic glycoside content are compared with preset thresholds to generate a test report containing a pass / fail judgment conclusion.

2. The detection method according to claim 1, characterized in that, In the adaptive baseline correction, the total training loss function of the physically constrained deep residual network is a weighted sum of the prediction loss and the physical constraint loss. The physical constraint loss is used to penalize cases where the additive bias vector of the network output deviates from the Kubelka-Munk scattering model and the temperature drift model. The initial value of the weight coefficient of the physical constraint loss is 0.

1. The spectral feature extraction backbone of the physically constrained deep residual network consists of four residual blocks connected in series. The physical constraint prior layer is a two-layer fully connected network with Softplus as the activation function. The intermediate features of the two branches are fused through the Hadamard product.

3. The detection method according to claim 1, characterized in that, The method for constructing the excipient spectral dictionary includes: selecting no fewer than 8 kinds of pure pharmaceutical excipients, repeatedly acquiring the spectrum of each pure excipient in the mid-infrared channel and the near-infrared channel no fewer than 10 times and averaging the results, normalizing the L2 norm of each excipient spectral vector, and concatenating all the normalized excipient spectral vectors column by column to form the mid-infrared excipient dictionary matrix and the near-infrared excipient dictionary matrix, respectively; the pure pharmaceutical excipients include microcrystalline cellulose, magnesium stearate, gelatin, cross-linked sodium carboxymethyl cellulose, fumed silica, hydroxypropyl methylcellulose, talc, and sucrose.

4. The detection method according to claim 1, characterized in that, In the objective function of the non-negative sparse coding optimization under L1 regularization constraints, the reconstruction error term is the square of the L2 norm of the difference between the corrected mixed spectral vector and the product of the auxiliary material dictionary matrix and the sparse coefficient vector. The L1 regularization coefficient is determined by 5-fold cross-validation, and the sparse coefficient vector satisfies the non-negative constraint. The optimization problem is solved iteratively using the coordinate descent method.

5. The detection method according to claim 1, characterized in that, The multivariate curve-resolved alternating least squares iterative decomposition uses the SIMPLISM pure variable detection algorithm to extract initial pure component spectral estimates. The iterative convergence criterion is that the rate of change of the Frobenius norm of the fitting residuals between two adjacent iterations is less than... Or the number of iterations exceeds 500.

6. The detection method according to claim 1, characterized in that, The joint component extraction criterion of the bidirectional orthogonal partial least squares method is to maximize the covariance between the joint score matrix of the mid-infrared channel and the joint score matrix of the near-infrared channel, and the NIPALS algorithm is used for iterative solution. Five-fold cross-validation is used to perform grid search on the number of joint components, the number of mid-infrared orthogonal components, and the number of near-infrared orthogonal components. The optimal parameter combination is determined with the minimum sum of the root mean square errors of the cross-validation of SDG content and cyanogenic glycoside content as the objective function. The search range of the number of joint components is 1 to 15, and the search range of the number of mid-infrared orthogonal components and the number of near-infrared orthogonal components is 0 to 5 respectively.

7. The detection method according to claim 1, characterized in that, The spectral acquisition resolution of the mid-infrared channel is 4 cm⁻¹. -¹ The cumulative number of scans was 32; the spectral acquisition resolution of the near-infrared channel was 1 nm, and the cumulative number of scans was 32; the raw spectral data of the dual channels were converted into a spectral matrix after absorbance conversion. The dual-channel spectral acquisition used an integrated spectrometer equipped with an attenuated total reflectance accessory to perform non-destructive in-situ scanning of the sample under test. For hard capsule samples, the outer shell did not need to be peeled off and the sample was placed directly on the probe plate for scanning. For compressed candies, the sample was placed in the diffuse reflectance integrating sphere accessory for scanning.

8. The detection method according to claim 1, characterized in that, The physically constrained deep residual network is trained using the Adam optimizer with an initial learning rate of [value missing]. The training rounds consist of 200 epochs with a batch size of 32, and an early stopping strategy is adopted where training stops when the loss on the validation set no longer decreases after 20 consecutive epochs. When the multivariate curve resolution-alternating least squares iterative decomposition is performed, the number of samples participating in the decomposition in this batch is no less than 20.

9. The detection method according to claim 1, characterized in that, The entire process, from the moment the sample is placed to the output of the quantitative test report, takes no more than 30 seconds, and the testing process does not require the consumption of chemical reagents.

10. The detection method according to claim 1, characterized in that, Also includes: After outputting the quantitative prediction value, the predicted value of cyanogenic glycoside content is compared with the preset safety threshold. When the predicted value of cyanogenic glycoside content exceeds the safety threshold, a safety warning is triggered and the batch is marked as unqualified. The predicted value of SDG content is compared with the label claim range. When the predicted value of SDG content deviates from the label claim range by more than 15%, the active ingredient content is marked as substandard.

Images